Moisture control method and system
By dynamically calculating the filtering parameters using a learning model, the adaptability problem of the moisture control system in the airflow drying machine was solved, achieving more efficient and stable moisture control and improving the quality of tobacco and energy utilization efficiency.
Patent Information
- Application Number
- CN202511843039.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-09
- Publication Date
- 2026-02-27
AI Technical Summary
In the existing technology, the moisture control system of the airflow drying machine relies on manually setting the filter parameters, which cannot adapt to changes in production conditions, resulting in poor control accuracy and stability, and serious energy waste.
The filtering parameters are dynamically calculated using a learning model. The mapping relationship of the filtering parameters is constructed through a random forest regression algorithm. Combined with real-time moisture value, standard deviation, variance, ambient temperature and humidity, and material grade, the filtering intensity is automatically adjusted to achieve adaptive filtering.
It improves the precision and stability of moisture control, reduces energy waste, enhances the uniformity and quality of tobacco drying, and adapts to changes in different production conditions.
Smart Images
Figure CN121569987A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of tobacco equipment control technology, and specifically to a moisture control method and system. Background Technology
[0002] The airflow drying machine is a key piece of equipment in cigarette manufacturing, which uses high-temperature process gases to rapidly dehydrate and shape the tobacco shreds. During this process, the stable and precise control of the moisture content (i.e., water content) of the exit material directly affects the filling value, sensory quality, and processing performance of subsequent steps.
[0003] Currently, the industry commonly uses a PID (Proportional-Integral-Derivative) control strategy to achieve closed-loop moisture control. Its basic principle is: the moisture content of the material is detected in real time by a moisture meter at the outlet, and this detected value is compared with a preset target value. The output power of the burner (i.e., the hot air temperature) is adjusted based on the deviation between the two values, thereby controlling the dehydration rate and ultimately stabilizing the outlet moisture content within the target range.
[0004] However, in actual production, it was found that because tobacco shreds are prone to clumping in high-speed airflow, the moisture meter detection signal contains non-process-related, severe instantaneous fluctuations (noise). If this unprocessed signal is directly fed into the PID controller, it will cause frequent and unnecessary adjustments to the burner power, which will not only exacerbate energy waste but also cause oscillations in the moisture control loop, seriously affecting control accuracy and stability.
[0005] To suppress such signal noise, existing technologies typically introduce a first-order hysteresis filtering algorithm at the front end of the control loop. This algorithm can effectively smooth instantaneous spikes, but its filtering effect heavily depends on a key parameter—the filtering coefficient α. A larger α value results in stronger filtering, but also greater system response hysteresis; a smaller α value results in faster response, but weaker anti-interference capability. Currently, the setting of this filtering coefficient α relies entirely on the operator's experience: engineers set a fixed value during system initialization based on the general characteristics of a specific brand. This manual setting method is difficult to quantify due to its experience and varies from person to person, making it impossible to guarantee optimality. When production conditions change, such as switching tobacco brands, fluctuations in ambient temperature and humidity, or differences in raw material batches, the fixed α value cannot be adaptively adjusted, often leading to a deterioration in filtering effect—either incomplete filtering causing control fluctuations, or over-filtering causing sluggish control response. Summary of the Invention
[0006] This application provides a moisture control method and system, which aims to solve the above-mentioned problems.
[0007] In one embodiment, a moisture control method is provided for controlling the moisture content of the material at the outlet of an airflow drying machine, comprising the following steps:
[0008] S1. Real-time collection of material moisture content at the outlet of the airflow drying machine;
[0009] S2. The moisture value of the material is filtered using a first-order hysteresis filter method to obtain the processed moisture value of the material.
[0010] The filtering parameters in the first-order hysteresis filtering method are obtained by learning the model.
[0011] S3. Obtain the difference between the processed material moisture value and the set material moisture value. If the difference is greater than the preset value, increase the output power of the burner; otherwise, decrease the output power of the burner.
[0012] The moisture content of the material is the percentage of moisture in the material.
[0013] The preset value is a range value, which takes the form of [-1, 1]. When the difference is within the preset value, the burner's output power adjustment stops.
[0014] In one scheme, in step S1, the moisture content of the material at the outlet of the airflow drying machine is collected at predetermined time intervals, and a sequence {y0, y1, ..., y} is defined sequentially. N}; where y0 is the current moisture content of the material at the outlet of the airflow drying machine; y1 is the moisture content of the material collected at the previous predetermined time; y N The moisture content of the material is collected N predetermined time intervals before y0;
[0015] The first-order hysteresis filtering method is as follows:
[0016]
[0017] Among them, Y N Y represents the moisture content of the material after processing. N-1 For Y N The moisture content of the material after the previous processing; α is the filtering parameter, α = (0, 1).
[0018] 3. The moisture control method according to claim 2, wherein the filtering parameters are obtained through the learning model:
[0019] α=f(X)
[0020] X = (S, σ, σ) 2 T env H env P)
[0021] Where X is the feature vector; S is the arithmetic mean of the material moisture values collected in real time; σ is the standard deviation of the material moisture values collected in real time, and σ0 is the standard deviation of the material moisture values collected in real time. 2T represents the variance of the material moisture content collected in real time. env This refers to the ambient temperature value; H env The value represents the ambient humidity; P indicates the grade value.
[0022] Specifically, the grade value (P) represents the classification code of the tobacco product specification. The ambient temperature and ambient humidity values are averages for each production batch.
[0023] In one embodiment, the learning model is constructed using a random forest regression algorithm; the feature vector includes a prediction set and a training set; the training set is input into the learning model to train the learning model; and the prediction set is input into the trained learning model, which outputs the filtering parameters in real time.
[0024] In one approach, the data for both the prediction set and the training set are derived from multiple consecutive data points of material moisture content, ambient temperature, and ambient humidity.
[0025] Specifically, the prediction set and training set data can also include corresponding historical data collected from different production batches.
[0026] The multiple consecutive material moisture values, ambient temperature values, and ambient humidity values represent data collected at multiple consecutive predetermined times, and the collected data are divided into a prediction set and a training set in a 1:1 ratio.
[0027] Specifically, the learning model building process is as follows:
[0028] The bootstrap sampling method constructs a subset of the dataset: From the prepared complete training set (e.g., the latest 5 batches of data), N samples are randomly selected to form a subset of the dataset, which is roughly the same size as the original training set. This process is repeated K times to generate K different subsets of the dataset.
[0029] Decision tree construction: For these K subsets of data, K decision trees are trained in parallel. When training each node of each decision tree, instead of using all features, a subset of features is randomly selected from all features, and then the optimal feature from this subset is chosen for splitting (e.g., using mean squared error as the splitting criterion). This step ensures the diversity of the trees.
[0030] Forest formation: After the above process is completed, a "forest" consisting of K decision trees is generated.
[0031] The model output (predicting the optimal α value) process is as follows:
[0032] When a new batch needs to predict α, the corresponding feature vector X is input into the trained random forest model.
[0033] Each tree makes independent predictions: Each decision tree in the forest will predict based on the new input vector X. new The eigenvalues are used to traverse from the root node to a leaf node according to the node splitting rule, and an independent predicted value α is output. i .
[0034] Ensemble prediction results: The random forest model arithmetically averages the prediction results of all decision trees and outputs the average as the final optimal dynamic filtering parameter α.
[0035]
[0036] The learning model established α=f(S,σ,σ 2 T env H env The mapping relationship between P and P.
[0037] Specifically, for different grades of products, referring to production experience, an optimal filtering parameter α is first manually set. Then, during the operation of the airflow drying machine, the learning model continuously optimizes and adjusts the filtering parameter α based on the production results, and continuously collects production data such as material moisture content, ambient temperature, and ambient humidity.
[0038] When the production data accumulates to 5 batches (a batch refers to a continuous production process of tobacco shreds with the same brand, formula, and process standards from raw material input to completion), the production data of these 5 batches is input into the learning model to train the learning model, obtain new optimal filtering parameters α, and then send them to the PLC for production using the new parameters.
[0039] Before each production run, the latest five batches of production data are input into the learning model to obtain the filter parameter α as the filter value for the production start-up phase. The first-order lag filter method is then used to filter the moisture content of the material, and the optimal dynamic filter parameter α is continuously updated to achieve automatic adjustment and optimization of the filtering algorithm.
[0040] Moisture fluctuation rate decreased by 25%, and the batch standard deviation and variance excellence rate increased from 70% to over 85% (the above data are based on actual measured data from Qujing Cigarette Factory).
[0041] In one embodiment, another aspect provides a moisture control system, comprising: a data acquisition module for real-time acquisition of the moisture value of the material at the outlet of an airflow drying machine; a feature calculation module for processing the moisture value of the material; a data storage module for storing the data; a learning module for acquiring filter parameters; and a control module for controlling the burner power; wherein the data acquisition module, the feature calculation module, the data storage module, and the learning module are all electrically connected to the control module.
[0042] The data acquisition module is also used to collect ambient temperature and humidity data from the airflow drying machine. The feature calculation module is also used to process the ambient temperature and humidity data, performing preprocessing on the raw data, including handling missing values (forward filling, linear interpolation, mean filling), ensuring timestamp continuity, verifying the reasonableness of the numerical range (moisture content: 10-16%), and calculating S, σ, and σ². 2 Value. Special
[0043] The feature calculation module is also used to transform the raw data into features (S, σ, σ) that can be used by the learning model. 2 T env H env P).
[0044] One embodiment also includes: a monitoring and evaluation module and a control and scheduling module; both the monitoring and evaluation module and the control and scheduling module are electrically connected to the control module.
[0045] Specifically, the monitoring and evaluation module includes: a performance monitoring submodule for monitoring models and system operating status, an effect evaluation submodule, an alarm management submodule for managing alarms and notifications, and a visualization submodule for providing a visual interface.
[0046] The effectiveness evaluation submodule is used to evaluate the model's effectiveness in actual production.
[0047] The control and scheduling module includes: a task scheduling submodule for scheduling various machine learning tasks, a process control submodule for controlling the execution of machine learning processes, a version management submodule for managing model and data versions, and a system configuration submodule for managing system configurations.
[0048] In one embodiment, the learning module includes a model management submodule and a prediction service submodule; the model management submodule is used to interface with the training set, and the prediction service submodule is used to interface with the prediction set.
[0049] The model management submodule includes: model training submodule, model validation submodule, and model optimization submodule.
[0050] Specifically, the model training submodule is used to execute the machine learning model training process; the model validation submodule is used to evaluate model performance and select the optimal model; and the model optimization submodule is used to automatically tune the model hyperparameters.
[0051] The prediction service submodule includes a feature input submodule, a model inference submodule, a result post-processing submodule, and an interface service submodule.
[0052] The feature input submodule is used to receive and process prediction requests; the model inference submodule is used to load the model and perform predictions; the result post-processing submodule is used to post-process the prediction results; and the interface service submodule is used to provide external calling interfaces.
[0053] The beneficial effects of this application are:
[0054] This model calculates the key parameter α in the first-order lag filtering algorithm online using a learning model (such as a machine learning model trained on historical data), abandoning the traditional method of relying on manual experience to set fixed values. The model can comprehensively consider multi-dimensional information such as real-time moisture content, its statistical characteristics (standard deviation, variance), environmental conditions, and material grade, dynamically outputting the current optimal filtering parameters, making the filtering process adaptive.
[0055] Because the filtering parameters can automatically optimize according to changes in production conditions, the system can more effectively filter out abnormal signal fluctuations caused by tobacco clumping, while maintaining a rapid response to actual process trends. This provides a more stable and reliable moisture feedback signal for subsequent PID control, thereby reducing the fluctuation range of outlet moisture.
[0056] Whether changing tobacco grades, responding to seasonal changes in environmental temperature and humidity, or handling differences in raw material characteristics between different batches, the learning model in this method can automatically adjust the filtering intensity to adapt to new operating conditions by analyzing the latest production data. This solves the inherent defect of fixed-parameter filters where performance degrades under changing conditions, giving the moisture control system greater versatility and robustness.
[0057] This invention delegates the adjustment of filter parameters to an automated model, reducing reliance on the experience of senior operators or process engineers, lowering operational risks and quality fluctuations caused by differences in human judgment, and facilitating the standardization and intelligentization of production control.
[0058] More stable moisture control avoids frequent and drastic adjustments in burner power caused by signal noise, thus helping to save energy. At the same time, more stable outlet moisture directly improves the uniformity of tobacco drying, which is of positive significance for ensuring the filling value, sensory quality, and subsequent processing performance of tobacco. Attached Figure Description
[0059] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0060] Figure 1This is a schematic diagram of a moisture control method in one embodiment of this application;
[0061] Figure 2 This is a control flowchart of the moisture control system in one embodiment of this application;
[0062] The module comprises: 1. Data acquisition module; 2. Feature calculation module; 3. Data storage module; 4. Learning module; 5. Control module; 6. Monitoring and evaluation module; 7. Control and scheduling module; and 8. Burner. Detailed Implementation
[0063] The specific embodiments of this application will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application. Similarly, the following examples are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0064] In this invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0065] This application proposes improvements and innovations, and presents the following embodiments.
[0066] In some implementations, please refer to Figures 1 to 2 One aspect provides a moisture control method for controlling the moisture content of the material at the outlet of an airflow drying machine, comprising the following steps:
[0067] S1. Real-time collection of material moisture content at the outlet of the airflow drying machine;
[0068] S2. The material moisture value is filtered using a first-order hysteresis filter method, and the processed material moisture value is obtained.
[0069] In the first-order lag filtering method, the filtering parameters are obtained through the learning model.
[0070] S3. Obtain the difference between the processed material moisture value and the set material moisture value. If the difference is greater than the preset value, increase the output power of the burner; otherwise, decrease the output power of the burner.
[0071] The moisture content of the material is the percentage of moisture in the material.
[0072] The preset value is a range value, which takes the form of [-1, 1]. When the difference is within the preset value, the burner's output power adjustment stops.
[0073] This model calculates the key parameter α in the first-order lag filtering algorithm online using a learning model (such as a machine learning model trained on historical data), abandoning the traditional method of relying on manual experience to set fixed values. The model can comprehensively consider multi-dimensional information such as real-time moisture content, its statistical characteristics (standard deviation, variance), environmental conditions, and material grade, dynamically outputting the current optimal filtering parameters, making the filtering process adaptive.
[0074] Because the filtering parameters can automatically optimize according to changes in production conditions, the system can more effectively filter out abnormal signal fluctuations caused by tobacco clumping, while maintaining a rapid response to actual process trends. This provides a more stable and reliable moisture feedback signal for subsequent PID control, thereby reducing the fluctuation range of outlet moisture.
[0075] Whether changing tobacco grades, responding to seasonal changes in environmental temperature and humidity, or handling differences in raw material characteristics between different batches, the learning model in this method can automatically adjust the filtering intensity to adapt to new operating conditions by analyzing the latest production data. This solves the inherent defect of fixed-parameter filters where performance degrades under changing conditions, giving the moisture control system greater versatility and robustness.
[0076] This invention delegates the adjustment of filter parameters to an automated model, reducing reliance on the experience of senior operators or process engineers, lowering operational risks and quality fluctuations caused by differences in human judgment, and facilitating the standardization and intelligentization of production control.
[0077] More stable moisture control avoids frequent and drastic adjustments in burner power caused by signal noise, thus helping to save energy. At the same time, more stable outlet moisture directly improves the uniformity of tobacco drying, which is of positive significance for ensuring the filling value, sensory quality, and subsequent processing performance of tobacco.
[0078] In one embodiment, in step S1, the moisture content of the material at the outlet of the airflow drying machine is collected at predetermined time intervals, and a sequence {y0, y1, ..., y} is defined sequentially. N}; where y0 is the current material moisture content at the outlet of the airflow drying machine; y1 is the material moisture content collected at the previous predetermined time; y N The moisture content of the material is collected N predetermined time intervals before y0;
[0079] The first-order lag filtering method is as follows:
[0080]
[0081] Among them, Y N Y represents the moisture content of the processed material. N-1 For Y N The moisture content of the material after the previous processing; α is the filter parameter, α = (0, 1).
[0082] The mathematical form (recursive formula) of the first-order lag filtering algorithm and the physical meaning of the filtering parameter α (weighting coefficient between 0 and 1) are clearly defined, so that the filtering process can be implemented accurately and without error according to this scheme.
[0083] The moisture content (S), standard deviation (σ), and variance (σ) are used to calculate the moisture content. 2 The environmental temperature and humidity (T_env, H_env), and grade (P) are collectively defined as the feature vector X that determines the filtering parameters. This design allows the learning model to focus not only on the "average level" (S) of moisture but also on its "fluctuation" (σ, σ_env). 2 It combines "environmental disturbances" (T_env, H_env) and "material properties" (P) to achieve a comprehensive and three-dimensional assessment of the production status, making the decision-making basis far more scientific and comprehensive than a single empirical value.
[0084] The filter parameters are obtained through learning the model:
[0085] α=f(X)
[0086] X = (S, σ, σ) 2 T env H env P)
[0087] Where X is the feature vector; S is the arithmetic mean of the real-time collected material moisture values; σ is the standard deviation of the real-time collected material moisture values, and σ0 is the standard deviation of the material moisture values. 2 T represents the variance of the material moisture content collected in real time. env This refers to the ambient temperature value; H env The value represents the ambient humidity; P indicates the grade value.
[0088] Specifically, the grade value (P) represents the classification code of the tobacco product specification. The ambient temperature and ambient humidity values are averages for each production batch.
[0089] The formula α=f(X) transforms the originally vague "experience of veteran workers" into a clear and quantifiable functional relationship, eliminating the subjectivity and inconsistency of manual adjustments and promoting the standardization of production process control.
[0090] In one embodiment, the learning model is constructed using a random forest regression algorithm; the feature vector includes a prediction set and a training set; the training set is input into the learning model to train it; and the prediction set is input into the trained learning model, which outputs real-time filtering parameters.
[0091] Random forests, by integrating a large number of decision trees and voting (taking the average), can effectively smooth the prediction error of a single tree, making the final output α value highly resistant to interference and stable, and particularly suitable for processing noisy data in industrial production.
[0092] The algorithm's two major randomnesses (autonomous sampling of data rows and random subset selection of features) ensure the distinctiveness and independence of each decision tree, preventing the model from rote memorizing random noise in the training data, thereby improving the model's generalization prediction ability for new batches and new working conditions.
[0093] In one embodiment, the data for both the prediction set and the training set are derived from multiple consecutive data points of material moisture content, ambient temperature, and ambient humidity.
[0094] Specifically, the prediction set and training set data can also include corresponding historical data collected from different production batches.
[0095] Multiple consecutive material moisture values, ambient temperature values, and ambient humidity values represent data collected at multiple consecutive predetermined times, with the collected data divided into a prediction set and a training set in a 1:1 ratio.
[0096] Specifically, the process of obtaining the filter parameter α using the learning model (random forest regression algorithm) is as follows:
[0097] The bootstrap sampling method constructs a subset of the dataset: From the prepared complete training set (e.g., the latest 5 batches of data), N samples are randomly selected to form a subset of the dataset, which is roughly the same size as the original training set. This process is repeated K times to generate K different subsets of the dataset.
[0098] Decision tree construction: For these K subsets of data, K decision trees are trained in parallel. When training each node of each decision tree, instead of using all features, a subset of features is randomly selected from all features, and then the optimal feature from this subset is chosen for splitting (e.g., using mean squared error as the splitting criterion). This step ensures the diversity of the trees.
[0099] Forest formation: After the above process is completed, a "forest" consisting of K decision trees is generated.
[0100] The model output (predicting the optimal α value) process is as follows:
[0101] When a new batch needs to predict α, the corresponding feature vector X is input into the trained random forest model.
[0102] Each tree makes independent predictions: Each decision tree in the forest will predict based on the new input vector X. new The eigenvalues are used to traverse from the root node to a leaf node according to the node splitting rule, and an independent predicted value α is output. i .
[0103] Ensemble prediction results: The random forest model arithmetically averages the prediction results of all decision trees and outputs the average as the final optimal dynamic filtering parameter α.
[0104]
[0105] The learning model established α=f(S,σ,σ 2 T env H env The mapping relationship between P and P.
[0106] Specifically, for different grades of products, referring to production experience, an optimal filtering parameter α is first manually set. Then, during the operation of the airflow drying machine, the learning model continuously optimizes and adjusts the filtering parameter α based on the production results, and continuously collects production data such as material moisture content, ambient temperature, and ambient humidity.
[0107] When the production data accumulates to 5 batches (a batch refers to a continuous production process of tobacco shreds with the same brand, formula, and process standards from raw material input to completion), the production data of these 5 batches is input into the learning model to train the learning model, obtain new optimal filtering parameters α, and then send them to the PLC for production using the new parameters.
[0108] Before each production run, the latest five batches of production data are input into the learning model to obtain the filter parameter α as the filter value for the production start-up phase. The first-order lag filter method is then used to filter the material moisture value, and the optimal dynamic filter parameter α is continuously updated to achieve automatic adjustment and optimization of the filtering algorithm.
[0109] Moisture fluctuation rate decreased by 25%, and the batch standard deviation and variance excellence rate increased from 70% to over 85% (the above data are based on actual measured data from Qujing Cigarette Factory).
[0110] Employing a "sliding window" mechanism, the system consistently uses the latest five batches of data for model training and parameter prediction, making it a perpetual "learner." It can automatically track equipment performance drift, changes in raw material characteristics, and seasonal environmental shifts, enabling dynamic and continuous optimization of filter parameters.
[0111] By retaining only recent data, the model's decisions are always based on the most relevant and up-to-date "work experience." This design allows the system to quickly forget outdated patterns, learn new ones, and respond very swiftly to changes in production conditions.
[0112] The five-batch data window size ensures that the model has a sufficient sample size to learn stable patterns during training, while also minimizing computational overhead and speed during retraining, allowing it to be completed within the production batch intervals and meeting the real-time requirements of industrial settings. Actual operational data (a 25% reduction in moisture fluctuation and an improvement in the control excellence rate to over 85%) validates the effectiveness of this strategy.
[0113] In one embodiment, another aspect provides a moisture control system, comprising: a data acquisition module 1 for real-time acquisition of the moisture value of the material at the outlet of the airflow drying machine; a feature calculation module 2 for processing the moisture value of the material; a data storage module 3 for storing the data; a learning module 4 for acquiring filtering parameters; and a control module 5 for controlling the burner power; the data acquisition module 1, feature calculation module 2, data storage module 3, and learning module 4 are all electrically connected to the control module 5.
[0114] Specifically, control module 5 is a PLC controller.
[0115] The data acquisition module 1 is also used to collect ambient temperature and humidity data of the airflow drying machine. The feature calculation module 2 is also used to process the ambient temperature and humidity data. The feature calculation module 2 is used to preprocess the raw data, including handling missing values (forward filling, linear interpolation, mean filling), ensuring timestamp continuity, verifying the reasonableness of the numerical range (moisture content: 10-16%), and calculating S, σ, and σ². 2 Values. Feature calculation module 2 is also used to convert the raw data into features (S, σ, σ) usable by the learning model. 2 T env H env P).
[0116] Specifically, the processing code for the levy calculation module is as follows:
[0117] Function: Transforms raw data into features usable by the model.
[0118]
[0119]
[0120]
[0121] By defining core components such as data acquisition module 1, feature calculation module 2, data storage module 3, learning module 4, and control module 5, and clarifying their connection relationships, a complete blueprint for an industrialized control system integrating software and hardware is outlined, which can directly guide system development and integration.
[0122] Each module has a clearly defined role (e.g., the feature calculation module is dedicated to data preprocessing), and they are electrically connected to form an efficient data and control flow. This modular design reduces system complexity, improves maintainability and scalability, and facilitates phased implementation and troubleshooting.
[0123] The system architecture fully covers the entire chain of "data perception → feature extraction → model calculation → parameter distribution → control execution", realizing a fully automatic closed loop from production site data to final control parameter adjustment, without the need for manual intervention, and achieving intelligent control.
[0124] In one embodiment, it further includes: a monitoring and evaluation module 6 and a control and scheduling module 7; both the monitoring and evaluation module 6 and the control and scheduling module 7 are electrically connected to the control module.
[0125] Specifically, the monitoring and evaluation module 6 includes: a performance monitoring submodule for monitoring models and system operating status, an effect evaluation submodule, an alarm management submodule for managing alarms and notifications, and a visualization submodule for providing a visual interface.
[0126] The effectiveness evaluation submodule is used to evaluate the model's effectiveness in actual production.
[0127] The control and scheduling module 7 includes: a task scheduling submodule for scheduling various machine learning tasks, a process control submodule for controlling the execution of machine learning processes, a version management submodule for managing model and data versions, and a system configuration submodule for managing system configurations.
[0128] In one embodiment, learning module 4 includes a model management submodule and a prediction service submodule; the model management submodule is used to interface with the training set, and the prediction service submodule is used to interface with the prediction set.
[0129] The model management submodule includes: model training submodule, model validation submodule, and model optimization submodule.
[0130] Specifically, the model training submodule is used to execute the machine learning model training process; the model validation submodule is used to evaluate model performance and select the optimal model; and the model optimization submodule is used to automatically tune the model hyperparameters.
[0131] The prediction service submodule includes a feature input submodule, a model inference submodule, a result post-processing submodule, and an interface service submodule.
[0132] The feature input submodule is used to receive and process prediction requests; the model inference submodule is used to load the model and perform predictions; the result post-processing submodule is used to post-process the prediction results; and the interface service submodule is used to provide external calling interfaces.
[0133] Specifically, the program for each module of the learning model is as follows:
[0134] I. Model Management Submodule
[0135] Model training submodule
[0136] Function: Executes the machine learning model training process.
[0137]
[0138]
[0139]
[0140] The model validation submodule evaluates model performance and selects the optimal model.
[0141]
[0142]
[0143] The model optimization submodule automatically tunes the model's hyperparameters.
[0144]
[0145] The model deployment submodule manages model versions and deployments.
[0146]
[0147]
[0148] II. Prediction Service Submodule
[0149] Feature input submodule function: Receive and process prediction requests.
[0150]
[0151]
[0152] The Model Inference submodule's function is to load the model and perform predictions.
[0153]
[0154]
[0155] The post-processing submodule performs post-processing on the prediction results.
[0156]
[0157]
[0158]
[0159]
[0160] Interface service submodule function: Provides external call interfaces.
[0161]
[0162]
[0163] III. Monitoring and Evaluation Submodule Performance monitoring submodule function: monitoring model and system operating status.
[0164]
[0165]
[0166] The effect evaluation submodule function is to evaluate the effectiveness of the model in actual production.
[0167]
[0168]
[0169] The alarm management submodule manages alarms and notifications.
[0170]
[0171]
[0172] Visualization submodule function: Provides a visual interface.
[0173]
[0174]
[0175] IV. Scheduling and Control Submodule
[0176] The task scheduling submodule is responsible for scheduling various machine learning tasks.
[0177]
[0178]
[0179] The process control submodule controls the execution of the machine learning process.
[0180]
[0181]
[0182] Version Management Submodule Function: Manages model and data versions.
[0183]
[0184]
[0185] The system configuration submodule's function is to manage system configuration.
[0186]
[0187]
[0188]
[0189] The added monitoring and evaluation module 6 enables the system to self-perceive its health status (performance monitoring), objectively evaluate control effectiveness (effect evaluation), proactively warn of anomalies (alarm management), and intuitively display its operational status (visualization), transforming "passive control" into "proactive management and control," and significantly improving the system's maintainability and reliability.
[0190] The control and scheduling module 7 is responsible for coordinating the pace of tasks such as model training, updating, and prediction, while the version management submodule ensures the traceability and safe rollback of model iterations. This enables complex machine learning processes to run stably, orderly, and in a controlled manner in industrial environments, avoiding production risks caused by task conflicts or version chaos.
[0191] Learning Module 4 is further divided into a model management submodule (training, validation, optimization) responsible for "internal development" and a prediction service submodule (input, inference, post-processing, interface) responsible for "external services." This cohesive design improves the specialization of each submodule, making model lifecycle management more standardized, prediction service response more efficient, and the overall system's intelligent core more robust and powerful.
[0192] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application. Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.
Claims
1. A moisture control method for controlling the moisture content of materials at the outlet of an airflow drying machine, characterized in that, Includes the following steps: S1. Real-time collection of material moisture content at the outlet of the airflow drying machine; S2. The moisture value of the material is filtered using a first-order hysteresis filter method to obtain the processed moisture value of the material. The filtering parameters in the first-order hysteresis filtering method are obtained by learning the model. S3. Obtain the difference between the processed material moisture value and the set material moisture value. If the difference is greater than the preset value, increase the output power of the burner; otherwise, decrease the output power of the burner.
2. The moisture control method according to claim 1, characterized in that, In step S1, the moisture content of the material at the outlet of the airflow drying machine is collected at predetermined time intervals, and the sequence {y0, y1, ..., y} is defined sequentially. N }; where y0 is the current moisture content of the material at the outlet of the airflow drying machine; y1 is the moisture content of the material collected at the previous predetermined time; y N The moisture content of the material is collected N predetermined time intervals before y0; The first-order hysteresis filtering method is as follows: Among them, Y N Y represents the moisture content of the material after processing. N-1 For Y N The moisture content of the material after the previous processing; α is the filtering parameter, α = (0, 1).
3. The moisture control method according to claim 2, characterized in that, The filtering parameters are obtained through the learning model: α=f(X) X=(S,σ,σ 2 ,T env ,H env ,P) Where X is the feature vector; S is the arithmetic mean of the material moisture values collected in real time; σ is the standard deviation of the material moisture values collected in real time, and σ0 is the standard deviation of the material moisture values collected in real time. 2 T represents the variance of the material moisture content collected in real time. env This refers to the ambient temperature value; H env The value represents the ambient humidity; P indicates the grade value.
4. The moisture control method according to claim 3, characterized in that, The learning model is constructed using a random forest regression algorithm; the feature vector includes a prediction set and a training set; the training set is input into the learning model to train the learning model; and the prediction set is input into the trained learning model, which outputs the filtering parameters in real time.
5. The moisture control method according to claim 4, characterized in that, The data in both the prediction set and the training set are derived from multiple consecutive data points of material moisture content, ambient temperature, and ambient humidity.
6. A moisture control system for executing the moisture control method as described in claim 5, characterized in that, include: The system includes a data acquisition module for real-time acquisition of the moisture content of the material at the outlet of the airflow drying machine, a feature calculation module for processing the moisture content of the material, a data storage module for storing the data, a learning module for acquiring filter parameters, and a control module for controlling the burner power; the data acquisition module, the feature calculation module, the data storage module, and the learning module are all electrically connected to the control module.
7. The moisture control system according to claim 6, characterized in that, Also includes: The system includes a monitoring and evaluation module and a control and scheduling module; both the monitoring and evaluation module and the control and scheduling module are electrically connected to the control module.
8. The moisture control method according to claim 7, characterized in that, The learning module includes a model management submodule and a prediction service submodule; the model management submodule is used to interface with the training set, and the prediction service submodule is used to interface with the prediction set.