Wine factory yeast room temperature and humidity intelligent environmental control system and method based on large model

Abstract

The invention relates to an intelligent environment control system and method for the temperature and humidity of a distiller's yeast room of a winery based on a large model, and belongs to the technical field of white spirit brewing automation. The method comprises the following steps: deploying an intelligent koji room management platform to collect and store temperature and humidity data, door and window state information, external environment and other information of a plurality of koji rooms in real time, and carrying out preprocessing and feature extraction on the data; constructing a temperature and humidity prediction model based on the features and historical data; establishing a dynamic temperature and humidity threshold interval and a multi-stage early warning mechanism in combination with starter propagation process experience; when the predicted value exceeds the limit, triggering a large model intelligent analysis module to analyze an override cause, and generating a regulation and control scheme; after manual auditing, an electric door and window executing mechanism is remotely controlled, and the executing effect is fed back to the model for iterative optimization. The system comprises an Internet of Things sensing layer, an edge computing gateway, a cloud platform and a man-machine interaction terminal. According to the method, expert experience precipitation is realized through an intelligent decision and feedback mechanism, and the consistency of finished koji quality is improved.

Description

Technical Field

[0001] The present invention relates to a large-scale model-based intelligent environmental control system and method for temperature and humidity in a winery koji room, belonging to the technical field of intelligent liquor brewing industry. Background Art

[0002] Traditional winery koji room temperature and humidity control relies on manual inspections and manual adjustment of doors and windows. This leads to three major pain points: delayed response to abnormalities, strong reliance on experience, and heavy labor load. Manual inspections are difficult to promptly respond to sudden temperature and humidity fluctuations, and nighttime control has blind spots. Technicians' subjective experience cannot guarantee consistent koji quality. Furthermore, frequent cross-room operations in high-temperature environments are physically demanding. With the transition from pile koji production to rack koji, the room temperature and product temperature are more closely intertwined. Traditional fixed-threshold alarm systems are no longer able to meet the needs of dynamic environmental control. There is an urgent need to develop an intelligent environmental control system with predictive capabilities. This system can predict temperature and humidity trends and intervene proactively, achieving a closed-loop management system that transitions from passive response to active control, ensuring a stable koji room environment and consistent product quality. Summary of the Invention

[0003] In order to solve the above problems, the present invention discloses a large-scale model-based intelligent environmental control system and method for temperature and humidity in the winery koji room. The specific technical solution is as follows: A method for intelligent environmental control of temperature and humidity in a winery koji room based on a large model, comprising the following steps: Step 1: Collect multi-source business data: real-time collection of temperature and humidity in the koji room, door and window status, koji room information, koji making strategy, koji making process, and external environment data; Step 2: Predict the future: Build a temperature and humidity prediction model and output the predicted values ​​for the future period; Step 3: Early Warning: Generate personalized baseline intervals and trigger over-control warnings; Step 4: Generate control plan: The large model analyzes the cause of the anomaly and generates a control plan; Step 5: Optimize the model: After manual confirmation, remotely control the actuator and use the feedback data to optimize the model.

[0004] Furthermore, the multi-source business data in step 1 specifically includes: Temperature and humidity of the koji room: product temperature, indoor temperature, indoor humidity; Door and window opening: The door and window opening can be adjusted in the range of 1-99%, and the door and window opening is graded as [1, 20%], (20%, 40%], (40%, 70%], (70%, 99%], including the front-open and back-close mode; Process stages: fermentation period, low temperature culture period, slow rise period, post-fire and moisture removal period; Turning movements: one turn, two turns; Block covering status: whether covered with sacks for insulation; External environmental data: outdoor temperature and humidity, wind speed, and rainfall.

[0005] Furthermore, before entering step 2, data preprocessing is performed. The specific preprocessing process includes: linear interpolation to repair missing data, threshold correction of outliers, and integration of the curve translation operation with meteorological data to construct derived features; Standardization of multi-source business data is achieved through a three-level data processing mechanism: When data is missing, short-term communication interruptions of ≤3 collection cycles are dynamically compensated using cubic spline interpolation. For long-term missing data, a replacement sequence is constructed based on historical data from the same period and annotated with confidence levels. A double verification mechanism is established by combining the 3σ statistical principle with the threshold range of the koji making process to filter abnormal jump data and trigger the sensor calibration process.

[0006] Furthermore, the key driving variables in the preprocessed multi-source data are identified, redundant features that have little contribution to temperature and humidity fluctuations are eliminated, and the temperature and humidity prediction model described in step 2 is constructed. The temperature and humidity prediction model is a fusion model based on the LASSO model + LightGBM model; The LASSO model is fed with 12 hours of historical time series data and dynamic features. The time series data includes temperature and humidity time series data, door and window openings, and meteorological parameters. The dynamic features include bending action, block covering and insulation, door and window openings, and equipment disturbances. After the LASSO model filters the features, the output of the LASSO model is used as the input of the LightGBM model. The LightGBM model corrects the residuals and outputs the temperature and humidity forecast curve for the next 8 hours. The specific process is as follows: The LASSO model is used for feature screening, and the objective function is: , Where: Represents the true value of the i-th sample; represents the feature vector of the i-th sample; β represents the feature weight vector; N represents the total number of temperature and humidity data points in the curved room; i represents the data at the i-th collection moment; λ represents the regularization strength hyperparameter; represents the weight of the jth feature; p represents the total number of all input features in the Qufang system; j represents the jth input feature in the Qufang system, and when j=1, it represents the "current room temperature" feature; The input features are the original features of temperature and humidity time series data, door and window openings, and meteorological parameters; the core features are screened through TimeSeriesSplit cross-validation; the basic prediction value and residual sequence at each moment are output; The feature subset filtered by the LASSO model is used as the input of the LightGBM model to calculate the residual , set the parameters of tree depth, number of leaves, learning rate, and feature sampling ratio; the prediction adopts weighted fusion, the formula is: ; Where: It represents the prediction error at time point t, that is, the difference between the true value and the value predicted by the LASSO model; t represents the specific time point of data collection; represents the actual collected value at time point t; Represents the prediction result of the LASSO model on the target variable at time point t; Represents the final prediction result after weighted fusion of LASSO and LightGBM models; represents the predicted value of the LASSO model; Represents the LightGBM model pair The predicted value of To avoid error accumulation, the model group is trained independently for each hour of the next 8 hours. The near-term prediction is based on the complete historical data, and the far-term prediction only uses the data from time t-12 to time t; Output the temperature and humidity forecast curve for the next 8 hours.

[0007] Furthermore, dynamic feature construction specifically includes: Turning disturbance coefficient: calculated based on the temperature / humidity change rate ratio 2 hours after turning; Ventilation efficiency index: indoor and outdoor temperature difference × door and window opening × wind speed; The process stages are coded in stages: fermentation stage = 1, low temperature culture stage = 2, slow rise stage = 3, post-fire and moisture removal stage = 4; Classify and encode the door and window opening: the door and window opening can be adjusted in the range of 1-99%, and the door and window opening is graded as [1,20%]=1, (20%,40%]=2, (40%,70%]=3, (70%,99%]=1, including the front open and back close mode. The continuous process parameters are normalized by Min-Max and mapped to the interval [0,1] to form a structured feature matrix that adapts to the characteristics of the koji making process.

[0008] Furthermore, in step 3, personalized temperature and humidity threshold intervals are generated based on the season, koji room location, koji material type, and process stage, and a three-level warning mechanism is established, namely: threshold edge warning Level 1, mild over-control adjustment Level 2, and severe over-control emergency intervention Level 3.

[0009] Furthermore, the large model construction in step 4 specifically includes: (1) Constructing a knowledge graph for koji production and environmental control: Retrieving recent process documents, equipment operation manuals, historical work orders, and temperature and humidity measurement records from the koji production workshop, and sorting out empirical rules through the scene restoration method; During data preprocessing, the BERT-BiLSTM-CRF model was used for domain entity recognition. A specialized vocabulary was constructed for the specific expressions of music-making scenarios, defining five major entity categories and their attributes: process stage, environmental parameters, equipment entity, fault type, and control action. The Stanford CoreNLP tool was used to parse complex sentences in process documents and extract causal rules with confidence levels. During the knowledge fusion phase, the DTW algorithm was used to align the timestamps of the zigzag turning operations with the temperature mutation curve. It was calculated that 87% of zigzag turning operations would cause a temperature rise of 0.8-1.2°C within 2 hours. This led to the construction of a spatiotemporal chain linking "zigzag turning event → temperature fluctuation → ventilation equipment control," and the ventilation efficiency index was quantified. The knowledge graph is designed with a four-layer topology: process stage layer, equipment entity layer, fault diagnosis layer, and control strategy layer. InfluxDB is also integrated to store equipment status time series data. (2) Design and analyze prompt words for analyzing the reasons for over-control of temperature and humidity in koji making, and guide the large model to perform reasoning in combination with the knowledge graph; (3) Output the root cause analysis of over-control and corresponding control suggestions.

[0010] Furthermore, the large model in step 4 analyzes the cause of the anomaly based on the knowledge graph of the zigzag and environmental control, and outputs a control plan sorted by energy consumption priority. The specific process is as follows: The control plan is evaluated according to the evaluation model based on the implementation effect: , Where: Indicates the absolute deviation between the actual temperature and the target temperature; Indicates the total energy consumption of the execution plan; Indicates the device health index; score indicates the scoring result; Successful solutions with a calculation result score > 0.8 are stored in the strategy library, and inefficient solutions with a score < 0.6 trigger model retraining. The generated solutions are sorted according to the score value.

[0011] The system for implementing the above-mentioned large-scale model-based intelligent environmental control method for temperature and humidity in the winery koji room includes: IoT perception layer: This includes temperature and humidity sensors, door and window status detection, and weather data collection modules. It is used to collect and integrate multi-source heterogeneous data from Qufang. Specifically: Sensor layer: product temperature probe, room temperature sensor, hygrometer, Equipment layer: electric door and window opening status, bending machine operation log, AGV operation track, Environmental layer: real-time data from weather stations, including temperature, humidity, wind speed, and rainfall. Process layer: labels of koji making process stages, koji material ratio parameters; Edge computing gateway: Time series data and dynamic feature preprocessing unit, used to implement data cleaning and enhancement, specifically: Missing values ​​​​are filled: linear interpolation and adjacent filling to repair missing values ​​​​of communication interruptions. Data correction: Correct the abnormal jump value of the sensor and eliminate the excessive data based on the threshold. Noise reduction processing: Moving average of temperature and humidity in the previous 6 hours to eliminate high-frequency noise of the sensor. Feature derivation: relationship between indoor and outdoor temperature difference and door and window opening, quantitative ventilation efficiency, temperature change rate within 2 hours after turning over, 24-hour moving average humidity; Cloud platform: integrates temperature and humidity prediction models, large model engines, and device control interfaces; Human-computer interaction terminal: provides early warning dashboard and plan review interface.

[0012] Furthermore, the arrangement process of the IoT perception layer is as follows: Arrange the electric door and window control system for the curved room; Multiple high-temperature resistant digital sensors are precisely embedded in the center of the curved block; Deploy multiple sets of integrated temperature and humidity sensors, along with oxygen and carbon dioxide sensors, in a sheltered location 1.5 meters above the ground. A 1:1 3D simulation model was built using ANSYS FLUENT software. The spatial layout of environmental control equipment was optimized through multi-physics field coupling analysis. Circulating fans, humidifiers, and cooling and heating systems were configured to optimize the air duct structure while ensuring the stability of the block production process. Set up small weather stations as public nodes for environmental monitoring; Build a complete IoT network. All sensor data is transmitted back to the edge gateway every 15 minutes via the LoRa wireless protocol, and is connected to the cloud platform in real time via the 4G network. It also integrates business of multi-source heterogeneous systems with the AGV scheduling system, MES production system, QMS quality management system, and meteorological data interface, forming a closed-loop system covering environmental perception, equipment control, and production collaboration.

[0013] The beneficial effects of the present invention are: This invention effectively balances the commonalities and individual characteristics of each koji room by constructing a dual-layer control architecture of "macro-framework + micro-model." While supporting autonomous control within each koji room, the system leverages the macro-framework to implement collaborative constraints and dynamic error correction, significantly enhancing the model's universality and ensuring consistent koji quality.

[0014] This invention deeply integrates IoT perception with intelligent algorithms, using a LASSO+LightGBM fusion model to accurately predict temperature and humidity for the next eight hours (with an error of ±1.5°C / ±2%RH). This model, combined with a dynamic threshold multi-level early warning mechanism, triggers intelligent decision-making within the large model. By integrating process experience data with a knowledge graph, a hierarchical optimization plan is generated, reducing the wear and tear associated with drastic adjustments to equipment such as doors and windows. Remote control from the management platform shortens the response cycle for exception handling. The koji-making process is digitally stored, transforming experience into knowledge and data into assets, gradually forming a reusable control knowledge base and improving the consistency of finished koji quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of the overall technical solution of the present invention; Figure 2 This is the distribution diagram of the temperature sensors on the shelves in the Qu room; Figure 3 This is a flow chart of the Qufang temperature and humidity prediction and early warning determination; Figure 4 It is the logic diagram of the large model decision module. DETAILED DESCRIPTION

[0016] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0017] The present invention builds a dynamic optimization and control system for the song-making environment through the deep integration of Internet of Things perception, time series prediction algorithm and large language model. Figure 1 As shown, the specific process when implementing the present invention is as follows: Step S1, deploy the koji room management platform on site, complete the data collection of temperature and humidity sensor equipment, door and window equipment, etc., the platform registers information such as koji making plan, process stage, process parameter adjustment during koji making, and pre-processes the data such as temperature and humidity, door and window status, etc. of the koji room.

[0018] In step S2, key driving variables are identified from the multi-source data preprocessed in step S1, redundant features that contribute little to temperature and humidity fluctuations are eliminated, and a temperature and humidity prediction model based on LASSO+LightGBM fusion is constructed. The historical 12-hour time series data and dynamic features are input, and the temperature and humidity prediction curve for the next 8 hours is output.

[0019] In step S3, personalized temperature and humidity threshold ranges are generated based on season, koji room location, koji material type, and process stage. A multi-level early warning mechanism is established, including threshold edge warning (Level 1), minor overshoot adjustments (Level 2), and severe overshoot emergency intervention (Level 3). When the predicted value exceeds the limit at Level 1, a warning message is sent to the koji room management platform. When the predicted value exceeds the limit at Level 2, a large-scale intelligent analysis model is triggered to output a fine-tuning solution. When the real-time value exceeds the limit at Level 3, an overshoot alarm is triggered on the platform, and automated, large-scale equipment is manipulated for rapid intervention to reduce energy consumption and damage to doors and windows caused by sudden opening and closing. Levels 1 and 2 use predicted values ​​for judgment, and deviations from the threshold are small. Level 3 monitors the actual value collected by the product temperature sensor, i.e., the real-time value. If the actual value deviates significantly from the standard, strong intervention is required.

[0020] In step S4, the large model combines the knowledge graph of curve and environmental control to analyze the cause of the anomaly, and generates a control plan based on historical operation records and environmental parameters (such as sorting by priority such as window opening angle, duration, energy consumption, etc.).

[0021] In step S5, after reviewing the control suggestion plan outputted in step S3, the Qufang management platform remotely controls and issues instructions to actuators such as electric doors and windows.

[0022] Step S6: Collect the temperature and humidity change data after execution, calculate the control effect score, and reversely train the prediction model and the large model knowledge base.

[0023] The following describes the execution process of each step in detail with reference to the embodiments: Step S1: The construction of the IoT perception layer focuses on the intelligent transformation of traditional qu rooms. Through a hierarchical implementation plan, 12 qu rooms are upgraded (including 2 full intelligent transformations and 10 basic sensor equipment installations): First, the electric door and window control system of the qu room is systematically transformed, and 21 high-temperature resistant digital sensors are precisely embedded in the center of the qu block (the three-dimensional distribution of the temperature probe is detailed in the table). Figure 2, the position will change when turning the song, and manual maintenance and correction on the platform are required). Simultaneously, 5 sets of integrated temperature and humidity sensors are deployed in a windproof position 1.5 meters above the ground, and 1 oxygen sensor and 1 carbon dioxide sensor are configured. Secondly, a 1:1 3D simulation model is constructed based on ANSYS FLUENT software. The spatial layout of environmental control equipment is optimized through multi-physics field coupling analysis. Circulating fans, humidifiers and cold and heat source systems are scientifically configured (an air source heat pump + air conditioning box combination solution is adopted in the pilot phase). The air duct structure is optimized while ensuring the stability of the song block production process. At the same time, a small meteorological station is added as a public node for environmental monitoring. Finally, a complete Internet of Things network is constructed. All sensor data is transmitted back to the edge gateway every 15 minutes through the LoRa wireless protocol, and is connected to the smart song room management platform in real time via the 4G network. It also realizes the business integration of multi-source heterogeneous systems with the AGV scheduling system, MES production system, QMS quality management system and meteorological data interface, forming an intelligent closed-loop system covering environmental perception, equipment control and production collaboration.

[0024] The Qufang Management Platform achieves standardization of multi-source business data through a three-level data processing mechanism: for data missing scenarios, the platform uses cubic spline interpolation to dynamically compensate for short-term communication interruptions (≤3 collection cycles), and for long-term missing data, it constructs an alternative sequence based on historical data of the same period and adds confidence annotations (0.6-0.9); then, it combines the 3σ statistical principle with the Qu-making process threshold range to establish a double verification mechanism, filter abnormal jump data, and trigger the sensor calibration process. In the dimension of feature construction, the disturbance coefficient of koji making (calculated based on the ratio of temperature / humidity change rate 2 hours after koji making) and the ventilation efficiency index (indoor and outdoor temperature difference × door and window opening × wind speed) were innovatively designed. At the same time, the process stages were coded in stages (fermentation period = 1, low-temperature culture period = 2, slow rise period = 3, post-fire and moisture removal period = 4), and the door and window openings were graded and coded as [1, 20%] = 1, (20%, 40%] = 2, (40%, 70%] = 3, (70%, 99%] = 1, including the front-open and back-close mode. The door and window openings have a dead zone and the adjustable range is 1-99%). Continuous process parameters such as wind speed were normalized by Min-Max and mapped to the [0, 1] interval to form a structured feature matrix adapted to the characteristics of the koji making process.

[0025] Step S2 uses the LASSO model to perform feature screening, and the objective function is: , The input features are the original 50-dimensional features (including time series, process, and environment categories). Through TimeSeriesSplit cross-validation, λ=0.1 is selected to screen out 15-dimensional core features (such as 6-hour lagged product temperature, door and window openings, wind speed, etc.). The basic prediction value and residual sequence at each moment are output.

[0026] The feature subset filtered by the LASSO model is used as the input of the LightGBM model to calculate the residual , set the parameters: tree depth = 5, number of leaves = 31, learning rate = 0.05, feature sampling ratio = 0.8; prediction uses weighted fusion, the formula is: .

[0027] To avoid error accumulation, the model group is trained independently for each hour in the next 8 hours. The near-term prediction is based on the complete historical data, and the far-term prediction only uses the data from t-12 to t.

[0028] Finally, the temperature and humidity for the next 8 hours are accurately predicted (error ±1.5℃ / ±2%RH).

[0029] like Figure 2 As shown, steps S2 and S3 form a closed loop for monitoring the temperature and humidity in the koji room. Its core environmental protection mechanism is built on the knowledge system of koji-making experts: by integrating expert experience and production data, a dynamic threshold decision rule library is established that integrates seasonal characteristics, koji room spatial layout, koji material physical properties and process stage divisions. Among them, personalized threshold ranges for product temperature, room temperature and humidity are configured as initial benchmarks on the management platform; as production data accumulates, the system extracts environmental control parameters based on high-quality production batches (physical and chemical index comprehensive score of 90 points or above) to construct a standard baseline, and continuously optimizes and iterates through a rolling time window algorithm, ultimately forming a golden reference curve for the koji-making environment. In order to reduce energy consumption and mechanical losses caused by frequent start-up and shutdown of equipment, while balancing control accuracy and system stability, a three-level progressive warning response mechanism is designed - threshold edge warning (Level 1), mild over-control adjustment (Level 2) and severe over-control emergency intervention (Level 3). Through a gradient control strategy, refined management that minimizes process disturbances is achieved. The three-level warning trigger and response mechanism is shown in Table 1 below:

[0030] When a Level 2 warning is triggered, the system executes steps S4-6 to start the decision-making process (see the complete business logic for details). Figure 3 ), the specific implementation process is as follows: In the engineering practice of building a knowledge graph for koji making at a winery, a 2024 intelligent transformation project at a winery has become a complete implementation example. The project team first retrieved nearly five years of process documentation from the koji workshop (including digital scans of handwritten process notes), 37 types of equipment operating manuals (such as a certain model's electric door and window commissioning guide), a total of 2,685 historical work orders (paper copies were recognized by optical character recognition and structured and stored in JSON format), and five years of temperature and humidity measurement records (paper copies were manually entered into a time-series database). The team then conducted two months of interviews with five senior koji makers and, using a scenario reconstruction method, identified 236 empirical rules (such as "During the rainy season, when humidity exceeds 85% RH for three consecutive hours, prioritize closing windows to 20% rather than activating high-energy dehumidifiers"). During data preprocessing, the BERT-BiLSTM-CRF model was used for domain entity recognition. A specialized vocabulary was constructed for specific expressions in koji-making scenarios (e.g., "second turn temperature control" corresponds to a process stage node). Five major entity categories and their attributes were defined: process stages (fermentation period, low-temperature incubation period, slow heating period, and post-fire dehumidification period), environmental parameters (product temperature, room temperature, and humidity), equipment entities (e.g., "4613311# Koji room rear skylight No. 1 OP15"), fault types (product temperature rise, CO2 accumulation), and control actions (primary turn temperature and humidity adjustment, covering with sacks for insulation, and opening or closing doors and windows). The Stanford CoreNLP tool was used to parse complex sentences in process documentation (e.g., "If the humidity retention exceeds the limit during the post-fire dehumidification period, the ventilation system must be fully activated and the koji turning must be delayed"), extracting 383 causal rules with confidence levels. During the knowledge fusion phase, the DTW algorithm was used to align the timestamps of 136 folding operations with the product temperature mutation curve. The calculation showed that 87% of folding operations would cause a temperature rise of 0.8-1.2°C within 2 hours (correlation coefficient > 0.76). A spatiotemporal correlation chain was constructed: "folding event → temperature fluctuation → ventilation equipment control." The ventilation efficiency index was quantified (a 10% increase in window angle reduces ventilation efficiency by 0.3°C / h). The knowledge graph was deployed in a Neo4j 4.4 Enterprise Edition cluster, with a four-layer topology: a process stage layer (including process parameter constraints), an equipment entity layer (with associated maintenance work orders), a fault diagnosis layer (root cause inference paths), and a control strategy layer (a library of historical success cases). InfluxDB was also integrated to store time-series data on device status (such as second-by-second curves of door and window openings).

[0031] Private deployment of DeepSeek-70B large model and deep coupling with knowledge graph, pre-set structured query template, when Level When an early warning is triggered, the system automatically generates a structured query such as "Current environment: 18th hour of the post-fire dehumidification period, product temperature 33.5°C (threshold 32±1.5°C), CO2 concentration 4820 ppm, associated with 4613311# skylight opening 45%. Please search for similar cases in the past three years based on the graph and output a control plan." The large model traverses the "high temperature-high CO2" association node in the knowledge graph, retrieves the overheating event in the koji room in August 2023, and combines it with the current door and window status time series data (opening 45% unchanged for 3 hours) to generate a three-level plan (for example: Plan 1: immediately close the windows to 30% and start the backup fan, with an estimated temperature drop of 1.2°C in 45 minutes; Plan 2: close the windows in stages, adjusting by 5% every 15 minutes, and reduce the temperature by 0.8°C / hour through natural ventilation; Plan 3: cover the koji pile and delay turning over, with a process impact coefficient of 0.3). The early warning prompt, root cause analysis, and control plan are pushed to the intelligent diagnosis function interface of the koji room management platform. The process engineer conducts manual review and decision-making based on the completed alarm information description, as follows: (1) Execute the recommended solution: After selecting Solution 1, the system remotely controls the skylight opening to 30%, starts the backup fan, monitors the product temperature change rate in real time (0.9°C / h), writes the actual temperature drop deviation ±0.1°C back to the map, and associates it with the current operation work order; (2) Parameter correction execution: If adjustment plan 2 is "close the window 10% every 15 minutes", and the humidity drops to 82%RH after execution, the system records the correction reason "avoiding CO2 accumulation caused by excessive window closing in a single time" and updates the control strategy weight (natural ventilation priority + 15%); (3) Experience plan entry: When the technician adopts an autonomous strategy (such as "covering with double-layer sacks + opening the west window") and successfully removes the warning, the system stores the operation steps (sack thickness 15 cm, window opening angle 25%) in a structured manner in the atlas strategy layer; (4) Handling of false alarms: Manually mark false alarm cases (such as false triggering caused by sensor drift) and synchronize the training data set to correct the prediction model.

[0032] A digital work order is generated for each operation, recording the operator, equipment response parameters (such as the opening of skylight No. 1 OP151 behind Qufang No. 1 in 4613311# from 70% to 40%) and execution effect (2-hour product temperature return threshold), and automatically linking to the original warning event.

[0033] The execution effect is based on the evaluation model: , High-energy-consuming equipment (such as cold and heat sources) will have its energy consumption weight appropriately increased. Successful solutions (score>0.8) will be stored in the strategy library, and inefficient solutions (score<0.6) will trigger model retraining.

[0034] The above implementation forms a closed loop of "empirical rules → data verification → model iteration", which helps to better train the model, drive accurate predictions and the continuous evolution of the knowledge network, and realize the valuable transformation of koji-making knowledge from experience to data. The abnormal response time is shortened by 82% (45 to 8 minutes), and the unmanned nighttime fault handling rate reaches 100%.

[0035] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless defined as such, will not be interpreted in an idealized or overly formal sense.

[0036] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A method for intelligent environmental control of temperature and humidity in a winery koji room based on a large model, characterized in that: The following steps are involved: Step 1: Collect multi-source business data: real-time collection of temperature and humidity in the koji room, door and window status, koji room information, koji making strategy, koji making process, and external environment data; Step 2: Predict the future: Build a temperature and humidity prediction model and output the predicted values ​​for the future period; Step 3: Early warning: Generate personalized baseline intervals and trigger over-control warnings; Step 4: Generate control plan: The large model analyzes the cause of the anomaly and generates a control plan; Step 5: Optimize the model: After manual confirmation, remotely control the actuator and use the feedback data to optimize the model.

2. The intelligent environmental control method for temperature and humidity in the winery koji room based on a large model according to claim 1 is characterized in that: The multi-source business data in step 1 specifically includes: Temperature and humidity of the koji room: product temperature, indoor temperature, indoor humidity; Door and window opening: The door and window opening can be adjusted in the range of 1-99%, and the door and window opening is graded as [1, 20%], (20%, 40%], (40%, 70%], (70%, 99%], including the front-open and back-close mode; Process stages: fermentation period, low temperature culture period, slow rise period, post-fire and moisture removal period; Turning movements: one turn, two turns; Block covering status: whether covered with sacks for insulation; External environmental data: outdoor temperature and humidity, wind speed, and rainfall.

3. The intelligent environmental control method for temperature and humidity in the winery koji room based on a large model according to claim 1 is characterized in that: Before entering step 2, data preprocessing is performed. The specific preprocessing process includes: linear interpolation to repair missing data, threshold correction of outliers, and integration of curve translation operation and meteorological data to construct derived features; Standardization of multi-source business data is achieved through a three-level data processing mechanism: When data is missing, short-term communication interruptions of ≤3 collection cycles are dynamically compensated using cubic spline interpolation. For long-term missing data, a replacement sequence is constructed based on historical data from the same period and annotated with confidence levels. A double verification mechanism is established by combining the 3σ statistical principle with the threshold range of the koji making process to filter abnormal jump data and trigger the sensor calibration process.

4. The intelligent environmental control method for temperature and humidity in the winery koji room based on a large model according to claim 1 is characterized in that: Identify key driving variables in the preprocessed multi-source data, eliminate redundant features that have little contribution to temperature and humidity fluctuations, and construct the temperature and humidity prediction model described in step 2. The temperature and humidity prediction model is a fusion model based on the LASSO model + LightGBM model; The LASSO model is fed with 12 hours of historical time series data and dynamic features. The time series data includes temperature and humidity time series data, door and window openings, and meteorological parameters. The dynamic features include bending action, block covering and insulation, door and window openings, and equipment disturbances. After the LASSO model filters the features, the output of the LASSO model is used as the input of the LightGBM model. The LightGBM model corrects the residuals and outputs the temperature and humidity forecast curve for the next 8 hours. The specific process is as follows: The LASSO model is used for feature screening, and the objective function is: ， Where: Represents the true value of the i-th sample; represents the feature vector of the i-th sample; β represents the feature weight vector; N represents the total number of temperature and humidity data points in the curved room; i represents the data at the i-th collection moment; λ represents the regularization strength hyperparameter; represents the weight of the jth feature; p represents the total number of all input features in the Qufang system; j represents the jth input feature in the curved room system, and when j=1, it represents the "current room temperature" feature; The input features are the original features of temperature and humidity time series data, door and window openings, and meteorological parameters. The core features are screened out through TimeSeriesSplit cross-validation. Output the basic prediction value and residual sequence at each moment; The feature subset filtered by the LASSO model is used as the input of the LightGBM model to calculate the residual , set the parameters of tree depth, number of leaves, learning rate, and feature sampling ratio; the prediction adopts weighted fusion, the formula is: ; Where: It represents the prediction error at time point t, that is, the difference between the true value and the value predicted by the LASSO model; t represents the specific time point of data collection; represents the actual collected value at time point t; Represents the prediction result of the LASSO model on the target variable at time point t; Represents the final prediction result after weighted fusion of LASSO and LightGBM models; represents the predicted value of the LASSO model; Represents the LightGBM model pair The predicted value of To avoid error accumulation, the model group is trained independently for each hour of the next 8 hours. The near-term prediction is based on the complete historical data, and the far-term prediction only uses the data from time t-12 to time t; Output the temperature and humidity forecast curve for the next 8 hours.

5. The intelligent environmental control method for temperature and humidity in the winery koji room based on a large model according to claim 4 is characterized in that: Dynamic feature construction specifically includes: Turning disturbance coefficient: calculated based on the temperature / humidity change rate ratio 2 hours after turning; Ventilation efficiency index: indoor and outdoor temperature difference × door and window opening × wind speed; The process stages are coded in stages: fermentation stage = 1, low temperature culture stage = 2, slow rise stage = 3, post-fire and moisture removal stage = 4; Classify and encode the door and window opening: the door and window opening can be adjusted in the range of 1-99%, and the door and window opening is graded as [1,20%]=1, (20%,40%]=2, (40%,70%]=3, (70%,99%]=1, including the front open and back close mode. The continuous process parameters are normalized by Min-Max and mapped to the interval [0,1] to form a structured feature matrix that adapts to the characteristics of the koji making process.

6. The intelligent environmental control method for temperature and humidity in the winery koji room based on a large model according to claim 1 is characterized in that: In step 3, personalized temperature and humidity threshold intervals are generated based on the season, koji room location, koji material type, and process stage, and a three-level warning mechanism is established, namely: threshold edge warning Level 1, mild over-control adjustment Level 2, and severe over-control emergency intervention Level 3.

7. The intelligent environmental control method for temperature and humidity in the winery koji room based on a large model according to claim 1 is characterized in that: The large model construction in step 4 specifically includes: (1) Constructing a knowledge graph for koji production and environmental control: Retrieving recent process documents, equipment operation manuals, historical work orders, and temperature and humidity measurement records from the koji production workshop, and sorting out empirical rules through the scene restoration method; During data preprocessing, the BERT-BiLSTM-CRF model was used for domain entity recognition. A specialized vocabulary was constructed for the specific expressions of music-making scenarios, defining five major entity categories and their attributes: process stage, environmental parameters, equipment entity, fault type, and control action. The Stanford CoreNLP tool was used to parse complex sentences in process documents and extract causal rules with confidence levels. During the knowledge fusion phase, the DTW algorithm was used to align the timestamps of the zigzag turning operations with the temperature mutation curve. It was calculated that 87% of zigzag turning operations would cause a temperature rise of 0.8-1.2°C within 2 hours. This established a spatiotemporal correlation chain from "zigzag turning event → temperature fluctuation → ventilation equipment control" and quantified the ventilation efficiency index. The knowledge graph is designed with a four-layer topology: process stage layer, equipment entity layer, fault diagnosis layer, and control strategy layer. InfluxDB is also integrated to store equipment status time series data. (2) Design and analyze prompt words for analyzing the reasons for over-control of temperature and humidity in koji making, and guide the large model to perform reasoning in combination with the knowledge graph; (3) Output the root cause analysis of over-control and corresponding control suggestions.

8. The method for intelligent environmental control of temperature and humidity in the winery koji room based on a large model according to claim 6 is characterized in that: The large model in step 4 analyzes the cause of the anomaly based on the knowledge graph of the zigzag and environmental control, and outputs a control plan sorted by energy consumption priority. The specific process is as follows: The control plan is evaluated according to the evaluation model based on the implementation effect: ， Where: Indicates the absolute deviation between the actual temperature and the target temperature; Indicates the total energy consumption of the execution plan; Indicates the device health index; score indicates the scoring result; Successful solutions with a calculation result score > 0.8 are stored in the strategy library, and inefficient solutions with a score < 0.6 trigger model retraining. The generated solutions are sorted according to the score value.

9. A system for implementing the large-scale model-based intelligent environmental control method for temperature and humidity in a winery koji room according to any one of claims 1 to 8, characterized in that: include: IoT perception layer: This includes temperature and humidity sensors, door and window status detection, and weather data collection modules. It is used to collect and integrate multi-source heterogeneous data from Qufang. Specifically: Sensor layer: product temperature probe, room temperature sensor, hygrometer, Equipment layer: electric door and window opening status, bending machine operation log, AGV operation track, Environmental layer: real-time data from weather stations, including temperature, humidity, wind speed, and rainfall. Process layer: labels of koji making process stages, koji material ratio parameters; Edge computing gateway: Time series data and dynamic feature preprocessing unit, used to implement data cleaning and enhancement, specifically: Missing values ​​​​are filled: linear interpolation and adjacent filling to repair missing values ​​​​of communication interruptions. Data correction: Correct the abnormal jump value of the sensor and eliminate the excessive data based on the threshold. Noise reduction processing: Moving average of temperature and humidity in the previous 6 hours to eliminate high-frequency noise of the sensor. Feature derivation: relationship between indoor and outdoor temperature difference and door and window opening, quantitative ventilation efficiency, temperature change rate within 2 hours after turning over, 24-hour moving average humidity; Cloud platform: integrates temperature and humidity prediction models, large model engines, and device control interfaces; Human-computer interaction terminal: provides early warning dashboard and plan review interface.

10. The large-scale model-based intelligent environmental control system for temperature and humidity in the winery koji room according to claim 9 is characterized in that: The arrangement process of the IoT perception layer is as follows: Arrange the electric door and window control system for the curved room; Multiple high-temperature resistant digital sensors are precisely embedded in the center of the curved block; Deploy multiple sets of integrated temperature and humidity sensors, along with oxygen and carbon dioxide sensors, in a sheltered location 1.5 meters above the ground. A 1:1 3D simulation model was built using ANSYS FLUENT software. The spatial layout of environmental control equipment was optimized through multi-physics field coupling analysis. Circulating fans, humidifiers, and cooling and heating systems were configured to optimize the air duct structure while ensuring the stability of the block production process. Set up small weather stations as public nodes for environmental monitoring; Build a complete IoT network. All sensor data is transmitted back to the edge gateway every 15 minutes via the LoRa wireless protocol, and is connected to the cloud platform in real time via the 4G network. It also integrates business of multi-source heterogeneous systems with the AGV scheduling system, MES production system, QMS quality management system, and meteorological data interface, forming a closed-loop system covering environmental perception, equipment control, and production collaboration.

Images