A method and system for predicting air conditioning start-up time suitable for a cigarette factory manufacturing process

Abstract

The application provides a kind of air conditioning start time prediction method and system suitable for cigarette factory manufacturing process, specifically includes system deployment and operation and maintenance, wherein the system deployment method includes, using the deployment scheme of background containerization, including the deployment of back-end service, the deployment of front-end interface and the deployment of front physical sensor, wherein the back-end service deployment includes deploying dedicated server.The advantage is that the data polling of appropriate time granularity is based on the feature engineering of multi-dimensional data to build a random forest to predict the air conditioning running time prediction, temporary adjustment strategy during prediction execution, and adaptive retraining of later model.The "trinity" air conditioning measurement and control effect of prediction intelligence, adaptive optimization of model and accurate measurement and control result is realized.

Description

Technical Field

[0001] This invention relates to a method and system for predicting air conditioner start-up time, and more particularly to a method and system for predicting air conditioner start-up time suitable for cigarette manufacturing processes, belonging to the field of intelligent equipment measurement and control. Background Technology

[0002] As a precision-manufactured product, cigarette production processes (such as tobacco processing, rolling, and packaging) place extremely stringent requirements on the temperature and humidity environment of the workshop. To ensure that the workshop environment is stable and meets the standards when production starts each day, the air conditioning system must be turned on in advance. Currently, the decision on start-up time mainly relies on the personal experience of operators. This traditional model has exposed three major pain points in actual operation: First, energy waste: To mitigate production risks, operators often set a long "safety buffer time," especially during seasonal transitions or when weather forecasts are inaccurate. Industrial air conditioning units are high-energy-consuming equipment, and prolonged unnecessary pre-operation accumulates into significant energy waste over time, contradicting management goals of "carbon neutrality" and cost reduction and efficiency improvement.

[0003] Second, production risk: Experience-based judgment ultimately has limitations. When facing atypical weather conditions (such as a sudden cold snap in spring or an unbearable heat wave in autumn), relying solely on intuition may lead to delayed start-up. If the workshop temperature and humidity are still below standard when production officially begins, it will directly affect the moisture content and physical properties of the tobacco, thus impacting key quality indicators such as the taste and combustibility of cigarettes. In severe cases, it may even necessitate suspending the production line until environmental standards are met, resulting in direct economic losses.

[0004] Third, management dependency: The system's operational efficiency and stability are highly dependent on the "feel" and "experience" of a few core employees, which is a fragile management model. When experienced employees are reassigned, take leave, or retire, knowledge gaps may occur, introducing uncertainty into the smooth operation of production. Datafication and modeling of this valuable experience is the only way to achieve knowledge accumulation and management modernization.

[0005] Fourth, existing technologies always strictly follow the predicted results in AI algorithms, without adapting the prediction results to the latest data to avoid mechanically following the predictions and ignoring the actual deviations in the prediction results.

[0006] Therefore, developing an intelligent model that can automatically sense environmental changes, especially the current performance based on predicted results, and accurately predict the time to achieve the target is a key technological measure to solve the above problems and promote the transformation and upgrading of factories towards "intelligent manufacturing". Summary of the Invention

[0007] 1. The core of this invention First, the deployment of physical sensors and SCADA data polling ensure that data can be queried periodically in the built-in InTouchHistorian database at a 1-minute time granularity, incrementally obtaining the latest values ​​of all relevant data points in the past minute, thus ensuring the accuracy and real-time performance of predictions.

[0008] Second, it collects multi-dimensional data covering environment, system conditions, and operating status, which can comprehensively reflect the accurate model of the system status.

[0009] Third, the core of the algorithm design is to abstract and transform the complex physical problem of "predicting the startup time" into a classic regression problem in supervised learning.

[0010] Fourth, we compared two industry-leading ensemble learning algorithms, XGBoost and Random Forest. Although their performance was similar, Random Forest was ultimately chosen because it demonstrated higher stability on the validation set of this project, was less sensitive to parameters, and was less prone to overfitting.

[0011] Fifth, to address model drift, we implemented a series of measures including performance monitoring, drift alerts, and the initiation of a pre-defined offline training process. This process involves retraining the model using the latest data, generating a new version of the model, and redeploying it, thus achieving adaptive optimization of the model.

[0012] Sixth, adopt a solution of temporarily changing the current air conditioning frequency by randomly sampling data.

[0013] 2. Specific technical solution of the present invention This invention provides a method for predicting air conditioning start-up time in cigarette manufacturing processes, including system deployment and maintenance. The system deployment method employs a containerized deployment scheme, encompassing backend service deployment, frontend interface deployment, and frontend physical sensor deployment. The backend service deployment includes deploying a dedicated server. The dedicated server includes an InTouch SCADA system that polls physical sensor data at a preset frequency, an InTouch Historian database, and a data acquisition service system. As a client, it periodically queries InTouch Historian with a 1-minute time granularity, incrementally retrieving the latest values ​​of all relevant data points within the past minute. The acquired data undergoes basic cleaning and format conversion before being stored in the Historian database. The dedicated server employs predetermined data feature engineering to characterize the data for model training and prediction. The training method includes offline training, and the prediction method involves online inference that incorporates randomly sampled data to change the air conditioner frequency. The operation and maintenance includes monitoring and retraining the model using the latest data based on model drift.

[0014] Optionally, the backend service deployment includes packaging the FastAPI application and all its dependencies (including model files) into a Docker image and deploying it on the dedicated server on the factory's internal network to run in a containerized manner; the frontend interface deployment involves building the React application into a set of static files (HTML / CSS / JS), hosted by the lightweight web server Nginx, and also deployed in a containerized manner.

[0015] The deployment of front-end physical sensors includes the deployment of physical sensors, including temperature sensors, humidity sensors, and pressure probes. The temperature sensors and humidity sensors include indoor temperature sensors and indoor humidity sensors, outdoor temperature sensors and outdoor humidity sensors, chilled water temperature sensors of the refrigeration system, and steam temperature sensors of the steam system. The pressure probes include chilled water supply pressure sensors of the refrigeration system.

[0016] Optionally, the data characterization includes constructing data features based on the initial outdoor temperature, initial indoor humidity, initial chilled water supply temperature, initial indoor temperature difference (ΔT) from the target temperature, initial indoor humidity difference (ΔH) from the target humidity, start-up time, initial moisture content, initial air enthalpy, and wet-bulb temperature.

[0017] Optionally, the models include XGBoost and Random Forest.

[0018] Preferably, the model includes a random forest.

[0019] Optionally, the offline training method includes the following steps: S1 data extraction, S2 data cleaning and labeling, S3 model training and validation S4 model evaluation.

[0020] Preferably, the offline training method further includes model persistence: serializing the trained and performance-compliant model object (containing all its learned parameters) into a file (e.g., random_forest_model_v2.joblib) and deploying it to a designated location on the production server for loading and calling by the online inference process.

[0021] Optionally, the online inference method that includes randomly sampling data to change the air conditioner frequency includes, for each qualified batch (i.e., products generated during the time period from the start of restarted production to the completion of production, emphasizing that only qualified batches are meaningful for change): P1 triggers the prediction task at regular intervals. P2 acquires current physical sensor data. P3 loads the model and constructs data features according to data featureization methods. P4 performs the prediction and outputs the results to the factory OPC UA server. When P5 adjusts the air conditioning to the preset time according to the prediction results, the OPC UA server polls the data nodes and the historical data of physical sensors in the InTouch Historian database. If the historical data does not match the data nodes, a temporary adjustment strategy is executed.

[0022] Optionally, the step of executing a temporary adjustment strategy when the historical data does not match the data node specifically includes the following steps: Q1. Obtain the time of arrival at the preset time. Previous batches of historical datasets Build historical data over time Change function , For a moment The time when the prediction task was previously triggered, among which For natural numbers, it represents the first natural number. The corresponding time of the batch The previous A preset time has arrived. In other words, historical data changes over time. The change function is a family of two-dimensional functions with N×M dimensions. Q2 Get the time of the current batch The previous preset time arrival time for this batch and time This batch of data and the time of this batch Data in data nodes ,Will By comparing historical data over time Change the function value and find the function value that is closest. Correspondence function of values , of which natural numbers and The result of the comparison is represented by time. The previous Batch No. The number of triggers, i.e. ,in Indicates time Pick The time range represents the corresponding function value. This indicates that the minimum value has been reached. Q3 Calculate target data ,judge Whether it is valid, If a threshold is set, and a match is found, the system runs according to the predicted duration; otherwise, if a mismatch occurs, the air conditioning frequency is changed, and the timing of the current batch is updated. Repeat steps Q2 and Q3 until the condition is met.

[0023] Optionally, the timer is set to 5-10 minutes, with a preset time of every 1-3 minutes. The historical data includes product temperature. ,humidity The method to change the air conditioner frequency is to follow the air conditioner frequency. and duration of action and product temperature humidity It is determined by the relational function, that is .

[0024] Specifically, it can be determined by historical air conditioning frequency. and duration of action and historical product temperature humidity The data was fitted to obtain the result. When a time interval is needed... Reach the target temperature and target humidity (i.e., timely) Use temperature symbols and humidity symbol express).

[0025] Monitoring and retraining the model involves three steps: performance monitoring, drift alerts, and one-click retraining. Optionally, the dedicated server includes a production server for controlling the production line and air conditioning, a factory OPC UA server for forming data nodes from the prediction results, a central monitoring system for the factory for subscribing to or polling the data nodes, a dedicated server for running the FastAPI application and all its dependent files, and a lightweight web server for hosting the files built with the React application.

[0026] The second objective of this invention is to provide a system capable of implementing the air conditioning start-up time prediction method suitable for cigarette factory manufacturing processes, comprising a back-end server, a front-end interface, front-end physical sensors, and an operation and maintenance system, wherein the back-end server includes a dedicated server.

[0027] 3. Beneficial effects The system employs a four-module deployment approach: backend service deployment, frontend interface deployment, frontend physical sensor deployment, and operation and maintenance deployment. This approach enables fine-grained polling and feature engineering of multi-dimensional data for training a superior random forest model. Temporary adjustment strategies are embedded during the prediction process, and drift alarms and model retraining based on model performance monitoring are used in later operation and maintenance. This achieves a "three-in-one" air conditioning measurement and control effect, which integrates intelligent prediction, adaptive model optimization, and accurate measurement and control results. Attached Figure Description

[0028] Figure 1 A general diagram of the four-module deployment scheme of the method of the present invention. Figure 2 A schematic diagram illustrating the data source and correlation comparison calculation method in the temporary adjustment strategy algorithm of this invention. Detailed Implementation

[0029] The implementation plan will explain the deployment of the method, the technology used for data interaction, the specific form of data, storage, the design and principle of the prediction algorithm, and system operation and maintenance.

[0030] 1. Method Deployment like Figure 1 This embodiment presents the deployment of four modules for a method suitable for predicting air conditioning start-up time in cigarette manufacturing processes. The first is system deployment, including backend service deployment, frontend interface deployment, and frontend physical sensor deployment. The second is operation and maintenance. The system deployment method employs a containerized deployment scheme for the backend, including backend service deployment, frontend interface deployment, and frontend physical sensor deployment. The backend service deployment includes deploying dedicated servers. To ensure high performance, high availability, maintainability, and future scalability, we selected a mature and well-established mainstream technology stack. Table 1 shows the overall deployment technology of the method. Table 1 Overall Deployment Technology of the Invention To ensure system stability and reliability, a containerized deployment solution is adopted: Backend service deployment: The FastAPI application and all its dependencies (including model files) are packaged into a Docker image. This image is deployed on a dedicated server on the factory's internal network, running as a container for easy isolation, migration, and scaling.

[0031] Front-end UI deployment: React applications are built as a set of static files (HTML / CSS / JS), hosted by the lightweight web server Nginx, and deployed in a containerized manner.

[0032] The deployment of front-end physical sensors, as shown in Table 1, includes the deployment of physical sensors, including temperature sensors, humidity sensors, and pressure probes. The temperature sensors and humidity sensors include indoor temperature sensors and indoor humidity sensors, outdoor temperature sensors and outdoor humidity sensors, chilled water temperature sensors of the refrigeration system, and steam temperature sensors of the steam system. The pressure probes include chilled water supply pressure sensors of the refrigeration system.

[0033] The dedicated servers involved in the method include a production server for controlling the production line and air conditioning, a factory OPC UA server for forming data nodes from the prediction results, a central monitoring system for the factory for subscribing to or polling the data nodes, a dedicated server for running the FastAPI application and all its dependent files, and a lightweight web server for hosting the files built into the React application.

[0034] 2. Technologies used for data interaction After system deployment and maintenance, the following details the data interaction methods: 1) Signal acquisition: Physical sensors (temperature, humidity, and pressure probes) installed at various locations on the air conditioning unit continuously measure physical signals.

[0035] 2) SCADA Polling: The factory's InTouch SCADA system polls these sensors at a preset frequency (e.g., on the order of seconds) to obtain raw readings.

[0036] 3) Primary storage: The SCADA system records the collected data points along with timestamps into its built-in InTouch Historian database.

[0037] 4) Data Synchronization: This project deployed an independent data acquisition service program. This service acts as a client, periodically initiating queries to InTouch Historian with a 1-minute time granularity to incrementally retrieve the latest values ​​of all relevant data points within the past minute.

[0038] 5) Secondary storage and preprocessing: The data acquisition service performs basic cleaning (such as removing invalid values) and format conversion on the acquired data, and then stores it in the Historian database dedicated to this project, providing a clean and independent data source for subsequent model training and online inference.

[0039] Data granularity: We chose 1 minute as the time resolution for data acquisition and storage. This granularity is fine enough to capture the dynamic process of temperature and humidity changes without putting unnecessary pressure on storage and computation due to excessive data volume.

[0040] 3. Specific format of the data Regarding the specific data format, in order to construct an accurate model that can comprehensively reflect the system state, we collected multi-dimensional data covering the environment, system operating conditions, and operational status, as shown in Table 2 below: Table 2 System Multidimensional Data

[0041] 4. Storage Regarding data storage, all collected data is stored in the Historian database in time-series format. Compared to traditional relational databases (such as MySQL), Historian, designed specifically for industrial process data, offers the following core advantages: 1. High-efficiency compression: It adopts advanced compression algorithms such as swing gate, which can store time-series data with extremely high compression ratio, saving a lot of disk space.

[0042] 2. Query performance: Deep optimization has been performed for time range queries, enabling fast retrieval of historical data within any time period.

[0043] 3. Data Aggregation: It has a rich set of built-in data aggregation functions (such as time-weighted average, maximum / minimum values, etc.), which facilitate downsampling and statistical calculations during data analysis and feature extraction.

[0044] 5. Design and Principles of Prediction Algorithms Regarding the design and principles of the prediction algorithm, the dedicated server employs predetermined data feature engineering to generate training and validation sets for model training and prediction.

[0045] Its core idea is that the core of the algorithm design is to abstract and transform the complex physical problem of "predicting the startup time" into a classic regression problem in supervised learning.

[0046] The logic is as follows: We assume the existence of an unknown mathematical function f that describes the relationship between the system's initial state and the final time to achieve the target. That is: Duration to achieve target = f(system state at boot time) Our task is to use a large amount of historical operational data and machine learning algorithms to find this function. An optimal approximate solution In the prediction process, we take the current real-time system state as input and substitute it into the obtained approximation function. This allows us to obtain a predicted timeframe for achieving the target. .

[0047]

[0048] in, It is a vector containing all relevant features at present. This is the model's predicted output (in minutes).

[0049] The training and validation sets used for model training are generated through feature engineering, or characterization, of the data. Data and features determine the upper limit of machine learning, while models and algorithms only approximate this limit. Feature engineering is the process of transforming raw data into features that can be better understood and utilized by the model, and it is crucial for project success. Table 3 shows the specific featureization content: Table 3. Data Characterization

[0050] After data feature generation, to address the selection and underlying principles of the algorithm, we compared and tested two industry-leading ensemble learning algorithms: XGBoost and Random Forest. Although their performance was similar, Random Forest was ultimately chosen due to its higher stability on the validation set in this project, its insensitivity to parameters, and its resistance to overfitting.

[0051] In-depth analysis of the principles of Random Forest: Random forest is an algorithm based on the Bagging ensemble concept and decision trees as base learners. Its power lies in the combination of "randomness" and "forest": Randomness manifests itself on two levels: Random sampling: When constructing each decision tree, instead of using all the training data, a subset of samples is obtained by random sampling with replacement from the total sample. This means that the learning content of each tree is slightly different.

[0052] Feature randomization: When splitting at each node of the decision tree, the optimal splitting feature is not selected from all features, but from a randomly selected subset of features.

[0053] Forest: The algorithm repeats the above process hundreds or thousands of times to build a "forest" consisting of a large number of independent and distinctive decision trees.

[0054] Ensemble Decision Making: When predicting new data, the data is simultaneously fed into every decision tree in the forest. Each tree independently provides a prediction. For regression problems (such as in this project), the final output of the random forest is the arithmetic mean of the predictions from all the decision trees.

[0055] This "three cobblers are equal to one Zhuge Liang" mechanism increases the diversity of the model by introducing randomness and eliminates individual bias through collective averaging, thus effectively avoiding the overfitting problem that is prone to occur with a single decision tree, making the model as a whole have extremely high accuracy and robustness.

[0056] Specifically, the entire system lifecycle is clearly divided into two independent yet interconnected closed-loop processes: offline training and online inference.

[0057] The offline training process is executed automatically in the background at a low frequency (e.g., weekly or monthly) to iteratively optimize the model using the latest running data, ensuring its predictive capabilities remain up-to-date. It includes the following steps: 1. Data Extraction: Query the Historian database for complete operating records of all air conditioners over the past year (or longer).

[0058] 2. Data cleaning and labeling: This is a key step in transforming raw logs into usable training samples.

[0059] a) Process identification: Write scripts to automatically identify each complete "boot-run-achieve-shutdown" process.

[0060] b) Definition of "Meeting the Standard": A precise definition of "meeting the standard" is required. This not only means that the temperature and humidity enter the target range, but also that after initially entering the target range, they can maintain a stable state for at least 5 minutes. This definition can effectively filter out false "meeting the standard" points caused by system oscillations.

[0061] c) Feature and label generation: For each identified valid process, extract the features (X) at its startup time and calculate the actual time consumed by the process (stable achievement timestamp - startup timestamp) as the learning target (y) of the model.

[0062] 3. Model Training and Validation: The processed dataset (X,y) is divided into a training set and a test set (e.g., 8:2).

[0063] The random forest model is trained on the training set using K-Fold Cross-Validation, and hyperparameters are tuned to find the optimal model configuration and obtain a reliable evaluation of the model's generalization ability.

[0064] 4. Model Evaluation: Evaluate the performance of the final model on independent test sets. Key evaluation metrics are shown in Table 4 below:

[0065] Regarding model persistence: The trained and performance-compliant model object (containing all its learned parameters) is serialized into a file (e.g., random_forest_model_v2.joblib) and deployed to a specified location on the production server for loading and use in the online inference process.

[0066] Regarding the online inference process, this is the core application of the system, running 24 / 7 and providing real-time prediction services. It includes an online inference method for randomly sampling data to change air conditioning frequencies. Specifically, for each qualified batch: P1 triggers a prediction task every 9 minutes, and a scheduled task scheduler (such as APScheduler) runs in the backend FastAPI service to trigger the prediction task once every 9 minutes.

[0067] P2 acquires current physical sensor data. After the task is triggered, the system immediately acquires "real-time snapshot" data of all relevant sensors of 16 air conditioning units through the data acquisition interface.

[0068] P3 loads the model and constructs features from the data using data featureization methods. The program loads the latest model file from disk into memory. Then, it performs the same feature engineering process as during training on the real-time acquired sensor data to construct feature vectors that meet the model input requirements.

[0069] P4 performs the prediction and outputs the results to the factory OPC UA server. Specifically, it inputs the constructed real-time feature vector into the loaded model, calls its predict() method, and instantly obtains an array containing 16 prediction durations (in minutes).

[0070] When P5 adjusts the air conditioning to the preset time according to the prediction results, the OPC UA server polls the data nodes and the historical data of physical sensors in the InTouch Historian database. If the historical data does not match the data nodes, a temporary adjustment strategy is executed. The backend service acts as an OPC UA client, connecting to the factory's OPC UA server. It writes the 16 predicted duration values ​​into 16 predefined data nodes (tags) on the OPC UA server. For example, 45.5 is written to the node ns=2;s=AC_System.Unit01.PredictedTimeToTarget. These data nodes are also used for polling.

[0071] Specifically, the provision that a temporary adjustment strategy is executed when the historical data does not match the data node includes, for example, the following: Figure 2 As shown, the polling interval is 1 minute, and the time from triggering the prediction task to the current time of this batch is given. (Other batches also correspond to the current time) The specific steps are as follows: (The text abruptly ends here, so the translation stops as well.) Q1 retrieves the time that arrives at preset intervals every 3 minutes. The previous M (Distinguished by upright and italic fonts) batches of historical datasets Build historical data over time Change function , For a moment The time when the prediction task was previously triggered, among which For natural numbers, it represents the first natural number. The corresponding time of the batch The previous The preset time has arrived.

[0072] That is, the independent variable time of the function of change is defined in, for example Figure 2To obtain a total of M batches, the trigger time of the graph for each batch is as follows: Between.

[0073] Q2 Get the time of the current batch (In this embodiment, the next trigger time) before the arrival time of the previous preset time for this batch. and time This batch of data and the time of this batch Data in data nodes ,Will By comparing historical data over time Change the function value and find the function value that is closest. Correspondence function of values , of which natural numbers and The result of the comparison is represented by time. The previous Batch No. The number of triggers, i.e. ,in Indicates time Pick The time range represents the corresponding function value. This indicates that the minimum value has been reached.

[0074] Among them, such as Figure 2 For each batch, with the prediction task triggered every 9 minutes, there are two preset arrival times within a prediction trigger cycle. There are 4 polling times for data nodes within each preset time period.

[0075] The "comparison" mentioned refers to comparisons between time periods in other historical batches outside of this batch. The function value is compared with each of the subsequent polling points.

[0076] Q3 Calculate target data ,judge Whether it is valid, If a threshold is set, and a match is found, the system runs according to the predicted duration; otherwise, if a mismatch occurs, the air conditioning frequency is changed, and the timing of the current batch is updated. Repeat steps Q2 and Q3 until the condition is met.

[0077] The historical data includes product temperature. ,humidity The method to change the air conditioner frequency is to follow the air conditioner frequency. and duration of action and product temperature humidity It is determined by the relational function, that is .

[0078] From a visualization perspective, the central monitoring system (HMI) on the factory's dedicated server acts as another OPC UA client, responsible for subscribing to or polling these data nodes. When an operator needs to power on the machine, the monitoring interface in the deployed front-end will directly display the model's predicted recommended duration and whether a temporary adjustment strategy has been implemented, thus achieving a closed loop from "experience-based decision-making" to "intelligent guidance."

[0079] 6. System Operation and Maintenance To address the risk of model drift (i.e., performance degradation due to factors such as equipment aging, process adjustments, and drastic seasonal changes), the system also includes an operation and maintenance mechanism: 1. Performance Monitoring: The system continuously records the model's predicted values ​​and the actual achievement values ​​afterward, and periodically calculates the true MAE and R. 2 .

[0080] 2. Drift Alert: Displays the trend of model performance changes on the operations and maintenance dashboard. When a performance metric (such as MAE) exceeds a preset threshold for a continuous period of time (e.g., MAE > 15 minutes), the system will automatically trigger an alert to notify the algorithm engineer.

[0081] 3. One-click retraining: After receiving an alert, engineers can initiate a preset offline training process, retrain the model using the latest data, generate a new version of the model, and redeploy it, thus forming a complete and continuously optimized MLOps (Machine Learning Operations) closed loop.

Claims

1. A method for predicting the start-up time of air conditioners in the manufacturing process of cigarette factories, characterized in that, This includes system deployment and maintenance. The system deployment method includes a containerized deployment scheme, encompassing backend service deployment, frontend interface deployment, and frontend physical sensor deployment. The backend service deployment includes deploying dedicated servers. The dedicated server includes an InTouch SCADA system that polls physical sensor data at a preset frequency, an InTouch Historian database, and a data acquisition service system. As a client, it periodically queries InTouch Historian with a 1-minute time granularity, incrementally retrieving the latest values ​​of all relevant data points within the past minute. The acquired data undergoes basic cleaning and format conversion before being stored in the Historian database. The dedicated server employs predetermined data feature engineering to characterize the data for model training and prediction. The training method includes offline training, and the prediction method involves online inference that incorporates randomly sampled data to change the air conditioner frequency. The operation and maintenance includes monitoring and retraining the model using the latest data based on model drift.

2. The method according to claim 1, characterized in that, The backend service deployment includes packaging the FastAPI application and all its dependent files into a Docker image and deploying it on the dedicated server on the factory's internal network, running it in container form; the frontend interface deployment involves building the React application into a set of static files, hosted by the lightweight web server Nginx, and also deploying it in a containerized manner.

3. The method according to claim 1, characterized in that, The deployment of front-end physical sensors includes the deployment of physical sensors, including temperature sensors, humidity sensors, and pressure probes. The temperature sensors and humidity sensors include indoor temperature sensors and indoor humidity sensors, outdoor temperature sensors and outdoor humidity sensors, chilled water temperature sensors of the refrigeration system, and steam temperature sensors of the steam system. The pressure probes include chilled water supply pressure sensors of the refrigeration system. The data characterization includes constructing data features based on the initial outdoor temperature, initial indoor humidity, initial chilled water supply temperature, initial indoor temperature difference (ΔT) from the target temperature, initial indoor humidity difference (ΔH) from the target humidity, start-up time, initial moisture content, initial air enthalpy, and wet-bulb temperature. Monitoring and retraining the model involves three steps: performance monitoring, drift alerts, and one-click retraining.

4. The method according to any one of claims 1-3, characterized in that, The models include XGBoost or Random Forest.

5. The method according to any one of claims 1-3, characterized in that, The model includes Random Forest.

6. The method according to claim 5, characterized in that, The offline training method includes the following steps: S1 data extraction, S2 data cleaning and labeling, S3 model training and validation S4 model evaluation; Online inference methods that involve randomly sampling data to change air conditioner frequencies include, for each eligible batch: P1 triggers the prediction task at regular intervals. P2 acquires current physical sensor data. P3 loads the model and constructs data features according to data featureization methods. P4 performs the prediction and outputs the results to the factory OPC UA server. When P5 adjusts the air conditioning to the preset time according to the prediction results, the OPC UA server polls the data nodes and the historical data of physical sensors in the InTouch Historian database. If the historical data does not match the data nodes, a temporary adjustment strategy is executed.

7. The method according to claim 6, characterized in that, The offline training method also includes model persistence: serializing the trained and performance-compliant model object into a file and deploying it to a specified location on the production server for loading and calling in the online inference process; The central monitoring system on the factory's dedicated server acts as another OPC UA client, responsible for subscribing to or polling these data nodes. When the operator needs to power on the machine, the monitoring interface in the deployed front-end interface will directly display the suggested duration of the model prediction. The dedicated servers include production servers that control the production line and air conditioning, factory OPC UA servers for forming data nodes from the forecast results, a factory central monitoring system for subscribing to or polling the data nodes, dedicated servers for running the FastAPI application and all its dependent files, and a lightweight web server for hosting the files built into the React application.

8. The method according to claim 7, characterized in that, When the historical data does not match the data node, the temporary adjustment strategy is executed, which specifically includes the following steps: Q1. Obtain the time of arrival at the preset time. Previous batches of historical datasets Build historical data over time Change function , For a moment The time when the prediction task was previously triggered, among which For natural numbers, it represents the first natural number. The corresponding time of the batch The previous The preset time has arrived. Q2 Get the time of the current batch The previous preset time arrival time for this batch and time This batch of data and the time of this batch Data in data nodes ,Will By comparing historical data over time Change the function value and find the function value that is closest. Correspondence function of values , of which natural numbers and The result of the comparison is represented by time. The previous Batch No. The number of triggers, i.e. ,in Indicates time Pick The time range represents the corresponding function value. This indicates that the minimum value has been reached. Q3 Calculate target data ,judge Whether it is valid, If the threshold is met, the system will operate according to the predicted duration; otherwise, the air conditioning frequency will be changed and the timing of the current batch will be updated. Repeat steps Q2 and Q3 until the condition is met.

9. The method according to any one of claims 6-8, characterized in that, The timing is 5-10 minutes, the preset time is every 1-3 minutes, and the historical data includes product temperature. ,humidity The method to change the air conditioner frequency is to follow the air conditioner frequency. and duration of action and product temperature humidity It is determined by the relational function, that is .

10. A system for predicting the start-up time of air conditioning in a cigarette manufacturing process, capable of implementing the method described in any one of claims 1-9, characterized in that, It includes a backend server, a frontend interface, frontend physical sensors, and an operation and maintenance system, wherein the backend server includes a dedicated server.

Images