Tobacco primary processing technology regulation and control method based on large model

Abstract

The invention discloses a tobacco primary processing technology regulation and control method based on a large model, and the method comprises the steps: carrying out the feature extraction at an edge layer according to technological parameters, raw material detection data and environmental parameters collected in a historical time interval, and obtaining a plurality of feature vectors; according to the multiple feature vectors, a quality index is obtained through prediction in an edge layer through a preset prediction model, and according to the quality index, process parameters are adjusted for execution; and according to the quality index acquired after execution and the quality index obtained by prediction, obtaining a prediction residual moving average at a decision-making layer, and when the prediction residual moving average is continuously greater than an error threshold for multiple times, triggering model re-calibration and synchronizing to an edge layer. According to the method, the multi-modal features are fused in the edge layer, closed-loop control is realized through the real-time prediction model, and the prediction accuracy and the response speed are improved. In combination with a prediction residual moving average dynamic calibration mechanism, the problem of model failure caused by parameter mutation is effectively solved, and production continuity is guaranteed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of manufacturing management prediction, and particularly relates to a tobacco cut tobacco process regulation method based on a large model. BACKGROUND

[0002] The tobacco cut tobacco process is a core link of cigarette production, and involves multiple processes such as cutting, drying and adding materials. Traditional process parameter adjustment relies on the experience of operators, and has the problem of strong dependence on manual experience. The operator manually adjusts parameters such as temperature and flow according to historical experience, which is difficult to respond to raw material batch fluctuations (such as water content and starch content differences) in real time, and has the problems of poor quality stability, high scrap rate and the like.

[0003] Therefore, statistical models such as PLS regression are gradually introduced to try to quantify the relationship between process parameters and quality indicators, and to realize automatic adjustment of parameters such as temperature and flow. Alternatively, a simple model (such as linear regression and decision tree) is trained based on historical data to predict indicators such as water content.

[0004] Although the prior art has tried to introduce automation and preliminary models, it has generally failed to effectively integrate process parameters, sensor data, raw material characteristics and other multi-source heterogeneous data, and the multi-modal data fusion is insufficient, resulting in limited model prediction accuracy.

[0005] In addition, when the model is arranged at the offline end, the offline end has limited computing power, and cannot recalibrate the model, or the model recalibration response speed is slow. When a new batch is put into production or the environment changes greatly, the model fails, the prediction result is unbalanced, and the process parameters are adjusted incorrectly, which further increases the waste rate of tobacco cut tobacco.

[0006] When the model is arranged at the cloud or database terminal, the collected parameters or preprocessed features need to be transmitted, and the process regulation parameter instructions are transmitted to the field equipment again. Due to the influence of the data volume and transmission speed of multi-source heterogeneous data, there is a problem of process regulation lag. Even in the offline case of network interruption, the model has the problem of failure.

[0007] Therefore, the prior art generally has the core problems of insufficient multi-modal data fusion, poor real-time performance and weak dynamic adaptability, and is difficult to cope with complex scenarios such as raw material batch fluctuations and environmental mutations, resulting in large quality fluctuation range and high scrap rate. It is urgent to develop a tobacco cut tobacco process regulation method based on a large model to comprehensively improve the stability and intelligent level of the cut tobacco process. SUMMARY

[0008] The application provides a tobacco primary processing process regulation method based on a large model, collects multi-source data of process parameters, raw material detection data and environmental parameters, solves the problem of insufficient model prediction accuracy, solves the problems of model recalibration lag and network failure through edge computing and online migration learning, and comprehensively improves the stability and intelligent level of the primary processing process.

[0009] The technical scheme adopted by the application is:

[0010] A tobacco primary processing process regulation method based on a large model comprises:

[0011] According to the process parameters, raw material detection data and environmental parameters collected in the historical time interval, feature extraction is performed in the edge layer to obtain a plurality of feature vectors;

[0012] According to the plurality of feature vectors, a quality index is predicted by a pre-set prediction model in the edge layer, and the process parameters are adjusted according to the quality index for execution;

[0013] According to the quality index collected after execution and the predicted quality index, a prediction residual moving average is obtained in the decision layer, when the prediction residual moving average is greater than an error threshold value for a plurality of times in succession, model recalibration is triggered, and the edge layer is synchronized.

[0014] The tobacco primary processing process regulation method based on a large model disclosed in the application also has the following additional technical features:

[0015] The process parameters, raw material detection data and environmental parameters are specifically:

[0016] The process parameters at least include any one of the temperature of a cut tobacco machine, steam flow, and the rotating speed of a cutter;

[0017] The raw material detection data at least include any one of the moisture content, starch content, total sugar, size and breakage rate of tobacco leaves;

[0018] The environmental parameters at least include any one of the environmental temperature and the environmental humidity.

[0019] Feature extraction is performed in the edge layer, and specifically:

[0020] According to the process parameters, raw material detection data and environmental parameters, corresponding statistical features, frequency domain features and time domain features are respectively processed;

[0021] According to the time stamp, the feature time of the process parameters, raw material detection data and environmental parameters is synchronized to obtain a plurality of feature vectors with fixed intervals in the historical time interval;

[0022] When the sampling time of the process parameters, raw material detection data, and / or environmental parameters does not coincide with the fixed interval time, the corresponding statistical features, frequency domain features, and time domain features are corrected by an interpolation method.

[0023] Before performing feature extraction on the edge layer, the method further comprises:

[0024] According to the process parameters, raw material detection data, and environmental parameters, abnormal values are filtered by a Gaussian distribution principle or a quartile range.

[0025] Missing values are detected by a sliding window method, and linear interpolation is used to fill in the missing values.

[0026] High-frequency noise is filtered by Kalman filtering or moving average algorithm smoothing processing.

[0027] In the edge layer, a quality index is predicted by a pre-set prediction model, specifically:

[0028] The plurality of feature vectors include a plurality of historical feature vectors and a real-time feature vector at the current time,

[0029] According to the plurality of historical feature vectors and the corresponding collected historical quality indexes, a prediction model is matched and mapped.

[0030] According to the matching and mapping logic, a quality index is predicted by the real-time feature vector.

[0031] The quality index at least includes any one of a moisture content of tobacco shreds and a temperature uniformity.

[0032] According to the quality index, the process parameters are adjusted for execution, specifically:

[0033] When the moisture content exceeds a first threshold value, and / or when the temperature uniformity is less than a second threshold value,

[0034] The process parameters are adjusted according to the quality index.

[0035] When the moving average of the prediction residual is greater than an error threshold value for a plurality of consecutive times, model recalibration is triggered, specifically:

[0036] When the moving average of the prediction residual is greater than an error threshold value for 5 consecutive times, model recalibration is triggered.

[0037] Model recalibration, specifically:

[0038] The Transformer bottom layer parameters of the model are frozen, and the control head parameters are adjusted by an L2 regularization loss function according to the historical data in the correction time interval before the model recalibration trigger.

[0039] Detect the distribution drift of the recalibration model using the isolation forest algorithm, and respond to the starting level according to the feature variation degree.

[0040] According to the feature variation degree, the starting level is responded, and specifically:

[0041] The feature variation degree includes first-level variation, second-level variation and third-level variation, and the variation degree gradually increases from the first-level variation to the third-level variation;

[0042] For the first-level variation, the collection frequency of the process parameters, the raw material detection data and the environmental parameters is increased;

[0043] For the second-level variation, at least part of the parameters in the Transformer bottom layer are triggered to be unfrozen, and the model calibration is re-executed;

[0044] For the third-level variation, the model version is rolled back to the previous version.

[0045] The application also provides a processing device, comprising:

[0046] A memory for storing a computer program;

[0047] A processor for executing the computer program to realize the steps of the tobacco primary processing process control method based on the large model.

[0048] Due to the adoption of the above technical solutions, the application has the following beneficial effects:

[0049] 1. In the application, the process parameters, raw material detection data and environmental parameters are extracted in the edge layer, solving the problem of insufficient multi-modal data fusion in traditional methods, providing multi-dimensional data input for subsequent prediction models, comprehensively considering the influence of multi-dimensional data on quality indicators, and improving the comprehensiveness and accuracy of model prediction.

[0050] The prediction model is deployed in the edge layer, the prediction work of the quality indicators is performed, and the process parameters are adjusted according to the prediction result. In this way, second-level closed-loop control is realized, the time consumption of data transmission is reduced, the response speed is significantly better than that of cloud computing, the production continuity is ensured, and model failure caused by network failure is avoided. According to the quality indicators, the effect is quantified, the quality defects are intercepted early, and the process parameters are adjusted to avoid excessive tobacco waste caused by single abnormality.

[0051] The model prediction accuracy is monitored in real time by calculating the moving average (MAE) of the prediction residual. When the MAE is greater than a threshold for a plurality of times in succession, the model is recalibrated to solve the problem of model failure caused by new batch production or environmental mutation. When the MAE is greater than the threshold for a plurality of times in succession, the model is recalibrated to avoid errors caused by occasional fluctuations in the environment and inaccurate single result prediction. When the MAE is greater than the threshold for a plurality of times in succession, it is ensured that the current raw material detection data and environmental parameters have changed and stabilized, and the model is recalibrated accordingly, thereby improving the stability of the model.

[0052] In addition, the decision layer is distinguished from the edge layer and is generally arranged in the cloud or a database terminal in a computer room, and has a much higher computing power than the edge layer. The model is recalibrated in the decision layer and is synchronized to the edge layer, greatly reducing the response time of model recalibration and improving the efficiency of process parameter adjustment in response to new batch production or environmental mutation.

[0053] In summary, the present application solves the core problems of poor quality stability, response lag and weak model adaptability in the traditional tobacco primary processing process through multi-source data fusion, real-time prediction and adjustment, and dynamic model calibration. BRIEF DESCRIPTION OF DRAWINGS

[0054] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0055] Figure 1 FIG. 1 is a flowchart of a tobacco primary processing process control method based on a large model according to an embodiment of the present application. DETAILED DESCRIPTION

[0056] In order to more clearly illustrate the overall concept of the present application, the following will be described in detail with reference to the accompanying drawings.

[0057] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be practiced in other manners different from those described herein, and therefore the scope of the present application is not limited by the specific embodiments disclosed below.

[0058] As shown in FIG. 1, a tobacco primary processing process control method based on a large model includes: Figure 1

[0059] S100: According to the process parameters, raw material detection data and environmental parameters collected in the historical time interval, feature extraction is performed in the edge layer to obtain a plurality of feature vectors.

[0060] ​The core purpose of this step is to provide high-quality input for subsequent prediction models by feature extraction and unified representation of multi-source heterogeneous data, solving the problem of insufficient multi-modal data fusion in traditional methods.

[0061] Multi-source data acquisition, real-time acquisition of process parameters (PLC interface signals), raw material detection data (NIR spectrum, visual detection), environmental parameters (temperature and humidity sensors).

[0062] Among them, the process parameters refer to the key parameters of the equipment operation in the cigarette making process, such as the temperature of the cut tobacco machine. If the temperature is too high, it will cause the cut tobacco to caramelize, destroy the aroma components, and affect the taste. If the temperature is too low, it will not effectively remove the moisture, resulting in unstable moisture content. By collecting process parameters in real time, the dynamic changes of the processing process can be directly reflected, providing key inputs for the quality prediction model.

[0063] Raw material detection data includes the moisture content, starch content, total sugar, size, breakage rate, and other physicochemical properties of tobacco leaves. The moisture content of raw materials directly affects the energy consumption and quality stability of the cut tobacco process. If the moisture content of raw materials is too high, the cut tobacco time needs to be increased, but too long may cause loss of aroma. The characteristics of raw materials determine the starting conditions of processing. Collecting raw material detection data can predict the processing difficulty in advance, avoid quality abnormalities caused by raw material fluctuations (such as tobacco from different origins), and reduce the scrap rate during batch switching.

[0064] Environmental parameters include environmental conditions such as temperature and humidity in the workshop. High temperature environment will accelerate the evaporation of cut tobacco moisture, causing the moisture content to deviate from the target value; low temperature environment may cause cut tobacco to absorb moisture, increasing the difficulty of subsequent cut tobacco. Environmental parameters are external interference factors of the processing process.

[0065] The collection of process parameters, raw material detection data, and environmental parameters is the basis for tobacco quality prediction and control. By fusing process parameters, raw material detection data, and environmental parameters, the dynamic correlation of the processing process (such as the nonlinear relationship between temperature and moisture content) can be fully captured, and the accuracy of the prediction model can be improved.

[0066] In addition, according to the collected process parameters, raw material detection data, and environmental parameters, statistical features, frequency domain and time domain features are extracted, and a fixed interval (such as 1 group per minute) feature vector sequence is generated, providing high-quality input for subsequent prediction models. Through the input of feature vectors, the dynamic correlation of process parameters (such as the nonlinear response of cut tobacco temperature and moisture content) can be easily captured, the complexity of prediction model calculation can be reduced, and the characteristic changes of parameters (such as sudden changes in starch content) can be quickly responded to, and timely adjustments of process parameters can be made.

[0067] In summary, this step solves the core problems of data heterogeneity, low prediction accuracy, and insufficient real-time performance in traditional cut tobacco process through innovative design of multi-source data fusion, feature extraction, and dimension reduction, laying a high-quality data foundation for subsequent model re-calibration and process control to achieve improved prediction accuracy of quality indicators, reduced scrap rate, and enhanced production continuity.

[0068] S200: According to the plurality of feature vectors, the quality indicators are predicted by the pre-set prediction model in the edge layer, and the process parameters are adjusted according to the quality indicators for execution.

[0069] The core purpose of this step is to realize closed-loop intelligent control of the tobacco cut process through real-time prediction and dynamic adjustment in the edge layer, solving the problems of large quality fluctuation and response lag caused by traditional manual experience control.

[0070] The trained prediction model, such as the TinyBERT model, is run in the edge layer to support real-time inference. It should be noted that a complete prediction model can be arranged in the edge layer, and a distilled prediction model (parameter amount < 10MB) can also be arranged to match the computing power of the edge layer to achieve fast prediction of quality indicators. The present application does not limit this.

[0071] Through the prediction model, the quality indicators at the current time are predicted according to the plurality of feature vectors. For cut tobacco, the quality indicators generally include moisture content and temperature uniformity, and temperature non-uniformity leads to tobacco coking or loss of aroma. According to the predicted quality indicators, the process parameters are adjusted and executed, such as ±2% opening of the tobacco dryer steam valve and ±5% speed of the tobacco cutter.

[0072] This step realizes minute-level parameter optimization through model prediction and automatic adjustment, improves adjustment speed, realizes second-level closed-loop control, responds to parameter changes in real time, reduces batch quality defects, and reduces cut tobacco scrap rate. In addition, the localized processing of the edge computing layer ensures uninterrupted production even in the case of network interruption.

[0073] In summary, this step solves the core problems of large quality fluctuation, response lag, and strong dependence on manual control in traditional cut tobacco process through real-time prediction and dynamic adjustment in the edge layer.

[0074] S300: According to the quality indicators collected after execution and the predicted quality indicators, a prediction residual moving average is obtained in the decision layer, and when the prediction residual moving average is greater than the error threshold value for multiple times in a row, model re-calibration is triggered and synchronized to the edge layer.

[0075] The core purpose of this step is to solve the model failure problem caused by new batch raw material production, environmental mutation or process parameter fluctuation through the model recalibration mechanism driven by prediction residual, and to improve the dynamic adaptability and robustness of the model.

[0076] According to the actual quality index actual_value (such as the moisture content at the cut tobacco outlet) collected after execution and the predicted quality index predicted_value (model output value), the absolute error is calculated,

[0077] residual = | actual_value - predicted_value |,

[0078] The moving average error (MAE) is obtained,

[0079] MAE_t = (1 / 5) * Σ (residual_i, i = t-4 to t).

[0080] The mean and standard deviation of MAE are calculated based on historical batch data, and the threshold is set to

[0081] T = μ_MAE + 2σ_MAE.

[0082] When MAE is greater than the threshold for multiple times in succession, it is determined that the model is invalid, the model failure is accurately identified, and misjudgment caused by accidental fluctuations (such as single temperature sensor noise) is avoided. The recalibrated model parameters are issued to the edge device (such as an industrial gateway) through the cloud. This step shortens the recalibration time to minutes through model calibration at the decision-making layer, and reduces the batch switching scrap rate.

[0083] This step solves the model failure problem caused by batch fluctuation of raw materials and environmental mutation in the traditional tobacco cut tobacco process through the model recalibration and edge synchronization mechanism driven by prediction residual.

[0084] As a preferred embodiment of the present application, the process parameters, raw material detection data and environmental parameters are specifically:

[0085] The process parameters at least include any one of the temperature of the cut tobacco dryer, the steam flow, and the rotating speed of the cutter;

[0086] The raw material detection data at least includes any one of the moisture content, starch content, total sugar, size, and breakage rate of the tobacco leaf;

[0087] The environmental parameters at least include any one of the environmental temperature and the environmental humidity.

[0088] The core purpose of the embodiment is to solve the problems of low model prediction accuracy and poor quality stability caused by incomplete parameter selection and insufficient data fusion in the traditional tobacco primary processing process through fine definition and combination of multi-source key parameters.

[0089] Among them, the process parameters refer to the key parameters of the equipment operation in the primary processing process. The cut tobacco drying machine temperature is collected in real time by a PT100 temperature sensor with an error of ±0.5%, which directly affects the drying effect of the cut tobacco. The steam flow is monitored by a mass flow meter (error ±0.5%) to control the uniformity of the added material (such as leaf moistening agent). The speed of the cutter: the speed change is indirectly reflected by a laser range finder (cutting width detection) to ensure the consistency of the cutting width.

[0090] The data sampling frequency is 1 time / minute, which is uploaded to the edge computing layer through the Modbus TCP / IP protocol.

[0091] The raw material detection data includes the moisture content, starch content, total sugar, size, breakage rate and other physicochemical properties of tobacco leaves. The moisture content is detected in real time by an online near-infrared spectrometer (NIR) with a sampling frequency of 1 time / minute. The starch content and total sugar are non-destructively and rapidly detected by NIR spectral analysis combined with chemometric algorithms. The size and breakage rate are identified by a visual detection system (CCD camera + OpenCV algorithm) to recognize the morphological characteristics of the leaves.

[0092] The environmental parameters include environmental conditions such as temperature and humidity in the workshop. The workshop environment is monitored by a temperature and humidity sensor (error ±1%) to compensate for the differences in cut tobacco hygroscopicity caused by changes in temperature and humidity. The data sampling frequency is 1 time / minute, which is uploaded to the edge computing layer through the MQTT protocol.

[0093] It can be understood that the traditional method only focuses on a single parameter (such as cut tobacco temperature) and ignores the synergistic effect of raw material characteristics (such as starch content) and environmental factors (such as temperature and humidity). The embodiment combines process parameters, raw material detection data and environmental parameters to construct multi-dimensional input features and more comprehensive feature representations, thereby improving the modeling capability of the model for complex process relationships.

[0094] The embodiment solves the core problems of incomplete parameter selection, low model prediction accuracy and weak dynamic adaptability in the traditional tobacco primary processing process through fine definition and multi-modal fusion of process parameters, raw material detection data and environmental parameters, thereby providing key data support for intelligent control of the tobacco primary processing process.

[0095] As a preferred embodiment of the present application, before performing feature extraction in the edge layer, the method further comprises:

[0096] According to the process parameters, raw material detection data and environmental parameters, abnormal values are filtered through the Gaussian distribution principle or quartile range.

[0097] Missing value detection by sliding window method to fill missing values by linear interpolation;

[0098] Smooth processing by Kalman filter or moving average algorithm to filter high-frequency noise.

[0099] The core purpose of the embodiment is to solve the problems of low model prediction accuracy and process parameter adjustment lag caused by poor data quality (such as abnormal values, missing values, and noise) in traditional tobacco cut tobacco process through data cleaning and preprocessing.

[0100] For abnormal value filtering, Gaussian distribution principle (3σ principle), assuming that the data follows normal distribution, removing data points deviating from the mean by more than 3 times the standard deviation (applicable to temperature, flow and other sensor data). Interquartile range method (IQR), calculate the interquartile range (IQR = Q3-Q1), remove outliers beyond Q1-1.5*IQR or Q3+1.5*IQR (applicable to non-normal distribution data, such as visual detection of leaf size). After abnormal value filtering, the data error rate is reduced, avoiding model overfitting to false data (such as temperature sensor false alarm).

[0101] For missing value filling, sliding window method detects missing values in a fixed time window (such as 5 minutes), and if the missing data rate in the window is >20%, the filling mechanism is triggered. Linear interpolation according to the mean of historical same conditions (such as the same raw material batch, similar environmental conditions) for linear interpolation. If linear interpolation is not available (such as new batch production), use the mean of historical same conditions (such as the same raw material characteristics) instead. After missing value filling, the data integrity is improved, avoiding model failure due to data missing.

[0102] For high-frequency noise smoothing, Kalman filter is suitable for dynamic systems (such as temperature changes of cut tobacco machine over time), which smooths noise through a prediction-correction mechanism. Moving average algorithm calculates the sliding window mean (such as 5-minute window) for time series data (such as steam flow) to suppress high-frequency jitter. After noise smoothing, the signal-to-noise ratio is improved, reducing prediction fluctuations.

[0103] The embodiment ensures the reliability of input data by multi-stage data cleaning (abnormal value filtering → missing value filling → noise smoothing), reduces the standard deviation of moisture content, and solves the problems of low model prediction accuracy and response lag caused by poor data quality in traditional tobacco cut tobacco process.

[0104] As a preferred embodiment under the present embodiment, feature extraction is performed in the edge layer, specifically:

[0105] According to the process parameters, raw material detection data, and environmental parameters, statistical features, frequency domain features, and time domain features are respectively processed.

[0106] According to the time stamp, the characteristic time of the process parameters, raw material detection data, and environmental parameters is synchronized to obtain a plurality of characteristic vectors with fixed intervals in the historical time interval;

[0107] When the sampling time of the process parameters, raw material detection data, and / or environmental parameters does not coincide with the time of fixed intervals, the corresponding statistical features, frequency domain features, and time domain features are corrected by an interpolation method.

[0108] The core purpose of the embodiment is to solve the problem of insufficient multi-modal data fusion in traditional methods by feature extraction and time synchronization of multi-source heterogeneous data, to provide high-quality input for subsequent prediction models, and to improve the comprehensiveness and accuracy of model prediction.

[0109] Statistical features are extracted, such as mean, variance, and slope, from process parameter time series data (such as temperature and flow within 30 minutes). For example, the mean of the cut tobacco drying machine temperature reflects the overall heat input, and the variance reflects the stability of temperature fluctuations.

[0110] Frequency domain features are extracted, such as dominant frequency components (such as the relationship between the vibration frequency of the cut tobacco drying machine and temperature fluctuations) through fast Fourier transform (FFT). For example, the frequency domain features of steam flow can reveal the influence of periodic fluctuations on uniform feeding.

[0111] Time domain features are extracted, such as peak amplitude (such as steam pressure mutation) and zero-crossing rate (such as periodic changes in feeding flow). For example, the zero-crossing rate of the cut tobacco drying machine speed reflects the periodic stability of the cutter roller operation.

[0112] IEEE 1588 protocol or hardware clock synchronization is used to unify multi-source data (such as PLC, NIR spectrum, temperature and humidity sensor) to the same time reference. For example, the sampling time of the cut tobacco drying machine temperature (15:00:00.123) and the sampling time of the NIR spectrum (15:00:00.145) are aligned to the same time point (15:00:00.130) through timestamp calibration.

[0113] The synchronized data is generated into a characteristic vector sequence according to fixed intervals (such as 1 group per minute), ensuring the continuity of model input.

[0114] It should be noted that the sampling frequencies of process parameters, raw material detection data, and environmental parameters may not be consistent. When there is no data at the fixed interval time point (15:00:00.130), linear interpolation is performed according to the characteristic vectors of the previous time and the next time to generate a replacement value, ensuring data integrity.

[0115] It can be understood that the space-time asynchrony of multi-source data can cause the loss of feature correlation (such as the asynchrony of temperature sensors and NIR spectrum sampling). The timestamp synchronization module (error <10 ms) ensures data consistency, provides reliable input for subsequent prediction models, reduces the feature vector deviation rate, provides comprehensive input for the prediction model, and reduces the prediction error caused by data defects.

[0116] The embodiment solves the core problems of insufficient data fusion, space-time inconsistency and integrity loss in the traditional tobacco cutting process through feature extraction, time synchronization and interpolation correction of multi-source heterogeneous data.

[0117] As a preferred embodiment of the present application, the quality index is predicted by a preset prediction model in the edge layer, and specifically:

[0118] The plurality of feature vectors include a plurality of historical feature vectors and a real-time feature vector at the current time,

[0119] According to the plurality of historical feature vectors and the historical quality index collected correspondingly, a prediction model is matched and mapped;

[0120] The quality index is predicted according to the matching and mapping logic through the real-time feature vector;

[0121] The quality index at least includes any one of moisture content of tobacco and temperature uniformity.

[0122] The core purpose of the embodiment is to solve the problems of large quality fluctuation and response lag caused by traditional manual experience control through the matching and mapping logic of historical feature vectors and real-time feature vectors.

[0123] It should be noted that the historical time interval includes the current time and a time interval formed in the past 30 minutes. Of course, other time spans containing the current time can also be used, and the present application does not limit this.

[0124] The input data in the embodiment include historical feature vectors containing statistical features (mean, variance), frequency domain features (FFT energy distribution) and time domain features (peak value, zero crossing rate) extracted in the past 30 minutes, and the corresponding historical quality index includes cut tobacco outlet moisture content (target ±0.2%) and temperature uniformity (standard deviation <1℃).

[0125] And the real-time feature vector at the current time, that is, the feature vector formed by the statistics, frequency domain and time domain features of the current process parameters (such as the temperature of the cut tobacco machine), raw material detection data (such as the NIR spectrum moisture content), and environmental parameters (such as the workshop temperature and humidity).

[0126] The historical feature vector and the historical quality index are input into the model to obtain a matched mapping logic (such as a Transformer attention mechanism). The real-time feature vector is input into the model to generate a quality index prediction value through the matched mapping logic (such as a Transformer attention mechanism). For example, moisture content prediction, based on the current drying temperature, steam flow, and NIR spectrum data, predicts whether the outlet moisture content exceeds the threshold value (13.5%). Temperature uniformity prediction, through the time domain features of the drying drum temperature distribution, calculates whether the temperature standard deviation is lower than 90%.

[0127] The matching mapping logic trained by historical data in the embodiment converts artificial experience into model rules, reduces parameter adjustment time, reduces cut tobacco rejection rate, and saves costs. In addition, the combination of historical feature vectors and real-time feature vectors captures the dynamic correlation between process parameters and quality indexes (such as the nonlinear relationship between temperature and moisture content), constructs a more comprehensive mapping logic, and improves prediction accuracy.

[0128] The matching mapping logic of historical feature vectors and real-time feature vectors in the embodiment solves the core problems of large quality fluctuations, response lag, and strong artificial dependence in traditional tobacco cutting processes.

[0129] As a preferred embodiment under the embodiment, the process parameters are adjusted according to the quality index to be executed, specifically:

[0130] When the moisture content exceeds the first threshold value, and / or when the temperature uniformity is less than the second threshold value,

[0131] The process parameters are adjusted according to the quality index.

[0132] The core purpose of the embodiment is to solve the problems of large quality fluctuations and response lag caused by traditional artificial experience regulation through a threshold value triggering mechanism based on the quality index.

[0133] The first threshold value (moisture content) is set to 13.5% according to historical excellent production batch data (such as the best moisture content control batch). The second threshold value (temperature uniformity) is set to 90% according to industry standards and equipment performance. By quantifying the threshold value (such as moisture content 13.5%), automatic adjustment is realized, artificial experience is converted into model rules, parameter adjustment time is reduced, and human decision interference is avoided.

[0134] When the moisture content is greater than 13.5% or the temperature uniformity is less than 90%, independent adjustment is triggered. When both exceed the standard, the moisture content is adjusted first (because the moisture content has a greater impact on combustion performance). Through the dual-index linkage trigger of moisture content and temperature uniformity, the pertinence of the adjustment strategy is improved.

[0135] In the present application, the adjustable parameters include, cut tobacco temperature, fine-tuning cut tobacco steam valve opening ± 2%, adjusted every 10 seconds. Cut tobacco speed, adjust the cutter speed according to the cut tobacco width distribution ± 5%. The flow rate of the feeder, the amount of leaf moistening agent sprayed is dynamically adjusted by the PID algorithm (such as steam flow ± 10%).

[0136] It should be noted that, in the process of adjusting the process parameters, through the incremental PID algorithm, the response lag of the actuator (such as the steam regulating valve) is eliminated, and system oscillation is avoided. The safety boundary constraint is executed, and the parameter adjustment range (such as temperature adjustment ± 5℃) is set to prevent equipment overload or process out of control.

[0137] The present embodiment solves the core problems of large quality fluctuation, response lag and strong manual dependence in the traditional tobacco cutting process through quality index threshold triggered process parameter adjustment.

[0138] As a preferred embodiment of the present application, when the moving average of the prediction residual is greater than the error threshold value for a plurality of consecutive times, the model is recalibrated, specifically:

[0139] When the moving average of the prediction residual is greater than the error threshold value for 5 consecutive times, the model is recalibrated.

[0140] The core purpose of the present embodiment is to solve the problem of model failure caused by raw material batch fluctuation, environmental mutation or process parameter adjustment lag through the trigger mechanism of moving average (MAE) of prediction residual exceeding the error threshold value for 5 consecutive times.

[0141] When the MAE calculated for 5 consecutive times is greater than the threshold value, it is determined that the model is invalid, and the model is recalibrated. For example, if the MAEs are 0.3%, 0.4%, 0.35%, 0.45% and 0.5%, and the threshold value is 0.3, the recalibration is triggered.

[0142] The present embodiment executes the 5-time trigger mechanism to avoid misjudgment caused by accidental fluctuations (such as single temperature sensor noise) and ensure the reliability of model failure determination.

[0143] As a preferred embodiment of the present embodiment, the model is recalibrated, specifically:

[0144] The Transformer bottom layer parameters of the model are frozen, and the control head parameters are adjusted by an L2 regularized loss function according to the historical data in the correction time interval when the model recalibration is triggered and before that;

[0145] The isolation forest algorithm is used to detect the distribution drift of the recalibrated model, and the corresponding start level response is started according to the feature variation degree.

[0146] The core purpose of the embodiment is to solve the model failure problem caused by data distribution change in the traditional model re-calibration process by combining the freezing of the Transformer bottom layer parameters and the control head fine-tuning, and the distribution drift detection driven by the isolation forest algorithm.

[0147] Freezing the Transformer bottom layer parameters retains the pre-trained Transformer encoder parameters (such as attention weights, position encoding), avoids overfitting of the bottom layer parameters due to changes in the distribution of new data, and improves the model update speed.

[0148] Fine-tune the control head parameters. When the model re-calibration is triggered, the historical data in the previous correction time interval (such as the past 30 minutes) is used as input to update the gradient of the control head (MLP layer), taking into account the stability and adaptability of the model. L2 regularization loss function is used Constrain the parameter update range to avoid overfitting caused by small sample data and improve the robustness of the model.

[0149] According to the isolation forest algorithm, the abnormal score s of the input features (such as starch content, tobacco leaf size) of the re-calibration model is calculated,

[0150]

[0151] where h(x) is the path length of the sample in the isolated tree, and c(n) is the expected path length for sample number n. According to the abnormal score, the degree of feature variation is quantitatively described, and the corresponding start level response is started.

[0152] The embodiment solves the model failure problem caused by data distribution change in the traditional tobacco primary processing process by combining the freezing of the Transformer bottom layer parameters and the control head fine-tuning, and the distribution drift detection driven by the isolation forest algorithm.

[0153] Specifically, according to the degree of feature variation, the corresponding start level response is as follows:

[0154] The degree of feature variation includes first-level variation, second-level variation and third-level variation, and the degree of variation gradually increases from first-level variation to third-level variation;

[0155] For first-level variation, increase the collection frequency of process parameters, raw material detection data, and environmental parameters;

[0156] For second-level variation, trigger the freezing of at least part of the Transformer bottom layer parameters to be removed, and re-execute the model calibration;

[0157] For third-level variation, start the model version rollback to the previous version.

[0158] The core purpose of the embodiment is to solve the model failure problem caused by data distribution mutation in the traditional model through the hierarchical response mechanism of the feature variation degree, and improve the dynamic adaptability and robustness of the model.

[0159] It should be noted that s≥0.9 is a third-level variation (high fluctuation), corresponding to a dramatic shift in feature distribution (such as a sudden change in starch content or sensor failure).

[0160] 0.7≤s<0.9 is a second-level variation (moderate fluctuation), corresponding to a moderate shift in feature distribution (such as a change in raw material batch).

[0161] 0.5≤s<0.7 is a first-level variation (low fluctuation), corresponding to a slight shift in feature distribution (such as a slow change in temperature and humidity).

[0162] In addition, it should be noted that s<0.5 is judged as no variation, and after model synchronization, the original data acquisition strategy is executed without the need for corresponding adjustment.

[0163] For first-level variation (low fluctuation), increase the sampling frequency, for example, increase the collection frequency of process parameters (such as drying machine temperature), raw material detection data (such as NIR spectrum moisture content), and environmental parameters (such as workshop temperature and humidity) from 1 time / minute to 1 time / 10 seconds, to ensure that new data fully covers the variation trend. For the synchronized model, only increase the sampling frequency to avoid triggering multiple recalibration due to slight fluctuation (such as slow change in temperature and humidity), and reduce the computational overhead.

[0164] For second-level variation (moderate fluctuation), partially unfreeze the Transformer underlying parameters, that is, unfreeze the attention weights of the Transformer encoder (retain 50% of the trained parameters), and combine L2 regularization (λ=0.001) to re-tune the control head, balancing model adaptability and stability.

[0165] For third-level variation (high fluctuation), roll back to the previous stable version of the model and trigger the manual review process to avoid long-term downtime and ensure production continuity.

[0166] The embodiment dynamically adjusts the response strategy according to the variation degree, optimizes resource allocation, and avoids resource waste or excessive intervention. By quantifying the severity of the anomaly score, the accuracy of the determination is improved.

[0167] The embodiment solves the model failure problem caused by data distribution mutation in the traditional tobacco cut tobacco process through the hierarchical response mechanism of the feature variation degree.

[0168] The application further provides a processing device, comprising:

[0169] a memory for storing a computer program;

[0170] A processor is configured to implement the steps of the large model-based tobacco primary processing method when executing the computer program.

[0171] Therefore, any effect of the large model-based tobacco primary processing method can be achieved, which is not described here.

[0172] The places not described in the present application can be realized by using or referring to the existing technology.

[0173] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other. Each embodiment mainly describes the difference from other embodiments.

[0174] The above only describes the embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the scope of claims of the present application.

Claims

1. A tobacco shred process control method based on a large model, characterized in that: include: Based on the process parameters, raw material testing data, and environmental parameters collected within the historical time interval, feature extraction is performed at the edge layer to obtain multiple feature vectors; According to the plurality of said feature vectors, a preset prediction model is used at an edge layer to predict a quality index, and according to the quality index, the process parameters are adjusted for execution; Based on the quality indicators collected after execution and the quality indicators obtained by prediction, the prediction residual moving average is obtained at the decision layer. When the prediction residual moving average is greater than the error threshold for multiple consecutive times, the model recalibration is triggered and synchronized to the edge layer.

2. The tobacco shred process control method based on a large model according to claim 1, characterized in that: Process parameters, raw material testing data, and environmental parameters, specifically: The process parameters include at least one of the temperature of the wire drying machine, the steam flow rate, and the speed of the wire cutting machine; The raw material detection data includes at least any one of the moisture content, starch content, total sugar, size, and breakage rate of the tobacco leaves; The environmental parameters include at least one of the environmental temperature and the environmental humidity.

3. The tobacco shred processing control method based on a large model according to claim 1, characterized in that: Feature extraction is performed at the edge layer, specifically: According to the process parameters, raw material test data, and environmental parameters, corresponding statistical features, frequency domain features, and time domain features are obtained respectively; Synchronize the characteristic time of the process parameters, raw material detection data, and environmental parameters according to the timestamp to obtain multiple characteristic vectors at fixed intervals within the historical time interval; When the sampling time of the process parameters, raw material detection data, and / or environmental parameters is inconsistent with the fixed interval time, the corresponding statistical characteristics, frequency domain characteristics, and time domain characteristics are corrected by an interpolation method.

4. The method for controlling tobacco shred processing based on a large model according to claim 3, characterized in that: Before feature extraction is performed at the edge layer, it also includes: According to the process parameters, raw material test data, and environmental parameters, outliers are filtered using the Gaussian distribution principle or interquartile range; Missing values ​​were detected by sliding window method and filled by linear interpolation; Smoothing is done through Kalman filtering or moving average algorithm to filter out high-frequency noise.

5. The tobacco shred process control method based on a large model according to claim 1, characterized in that: At the edge layer, the quality indicators are predicted through the preset prediction model, specifically: Multiple feature vectors include multiple historical feature vectors and real-time feature vectors at the current moment. Matching mapping logic through a prediction model based on the plurality of historical feature vectors and corresponding collected historical quality indicators; Predicting quality indicators based on the matching mapping logic using the real-time feature vector; Wherein, the quality index includes at least any one of the moisture content and temperature uniformity of the tobacco.

6. The method for controlling tobacco shred processing based on a large model according to claim 5, characterized in that: According to the quality indicators, the process parameters are adjusted for execution, specifically: When the moisture content exceeds a first threshold, and / or when the temperature uniformity is less than a second threshold, According to the quality index, the process parameters are adjusted.

7. The method for controlling tobacco shred processing based on a large model according to claim 1, characterized in that: When the moving average of the prediction residual is greater than the error threshold for multiple consecutive times, the model recalibration is triggered, specifically: When the moving average of the prediction residual is greater than the error threshold for five consecutive times, model recalibration is triggered.

8. The method for controlling tobacco shred processing based on a large model according to claim 7, characterized in that: Model recalibration, specifically: Freeze the underlying Transformer parameters of the model and adjust the control head parameters using an L2 regularized loss function based on historical data from the time when the model recalibration was triggered and the previous calibration time interval; The distribution drift of the recalibrated model was detected using the isolation forest algorithm, which corresponds to the activation level response according to the degree of feature variation.

9. The method for controlling tobacco shred processing based on a large model according to claim 8, characterized in that: According to the degree of feature variation, the corresponding activation level response is as follows: The degree of characteristic variation includes primary variation, secondary variation and tertiary variation. From primary variation to tertiary variation, the corresponding degree of variation gradually increases; For primary variation, increase the frequency of collecting process parameters, raw material testing data, and environmental parameters; For secondary mutation, at least some of the underlying parameters of the Transformer are unfrozen and the model calibration is re-executed; For level 3 mutations, initiate a model version rollback to the previous version.

10. A processing device, characterized in that: include: Memory for storing computer programs; A processor is used to implement the steps of the tobacco shred processing control method based on a large model as described in any one of claims 1 to 9 when executing the computer program.

Images