Digital factory intelligent control method and system based on AI
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-06
- Publication Date
- 2026-03-10
AI Technical Summary
Existing technologies are insufficient in feature mining and lack model collaboration in digital factories, making it difficult to adapt to rapidly changing market demands and complex production environments, and failing to fully realize the value of AI.
The method employs multi-source data acquisition, multi-level feature extraction, and multi-task AI model training. Using three levels of features as input, it trains four types of models: production scheduling optimization, quality prediction and parameter optimization, equipment fault early warning, and energy consumption optimization, thereby achieving real-time monitoring and closed-loop feedback, and incremental learning to update model parameters.
It achieves multi-objective collaborative decision-making, reduces instruction conflicts in traditional single-task models, improves production efficiency and energy consumption optimization, reduces manual intervention costs, and ensures long-term stability of control accuracy.
Smart Images
Figure CN121635142A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the technical field of intelligent control, and in particular to an AI-based digital factory intelligent control method and system. BACKGROUND
[0002] With the development of Industry 4.0, the concept of digital factory has gradually been accepted by the public. Traditional factory control systems rely on pre-programmed logic controllers (PLC) and manual intervention, and are difficult to adapt to rapidly changing market demands and complex production environments. In modern manufacturing, it realizes the digitalization and intelligent management of the production process with the help of advanced information technology and automation technology. The key to this change lies in the integration of various sensors, actuators and information systems to realize real-time monitoring and optimization of the production process. In recent years, some leading manufacturing enterprises have tried to introduce AI technology to optimize control, but due to the simplicity of feature extraction, single-task model and unreasonable deployment architecture, the value of AI has not been fully realized. SUMMARY
[0003] The purpose of the present application is to provide an AI-based digital factory intelligent control method and system, which can solve the problems of insufficient feature mining and lack of model cooperation in the prior art.
[0004] Technical solution: To solve the above technical problems, according to one aspect of the present application, more specifically, an AI-based digital factory intelligent control method, which specifically comprises the following steps: S1, multi-source data acquisition: collecting equipment operation, production process, quality detection, environment and energy consumption data; S2, multi-source data preprocessing: generating original feature data set after preprocessing the collected multi-source data by cleaning and normalizing; S3, three-level feature extraction: extracting features from the original feature data set according to data types to obtain first-level features, generating second-level features based on the management relationship and weighting of the first-level features, and obtaining third-level features based on the extraction of high-order semantics and prediction features from the second-level features; S4, multi-task AI model training: taking the third-level features as input, training four types of models including production scheduling optimization, quality prediction and parameter optimization, equipment fault early warning and energy consumption optimization; S5, real-time monitoring and execution: real-time acquisition of multi-source data, input of the preprocessed and three-level deep feature extraction into the AI model, generation and issuance of control instructions; S6, closed-loop feedback and model iteration: collecting execution results, evaluating the effectiveness of the third-level features, optimizing the feature extraction algorithm and incrementally updating the AI model parameters.
[0005] Furthermore, the multi-source data includes equipment operation data, production process data, quality detection data and environment and energy consumption data. Device data includes: device rotation speed, current, temperature, vibration frequency, fault code; Production process data includes: order quantity, production cycle, raw material ratio, process progress; Quality detection data includes: product size error, appearance defect, performance parameter; Environment and energy consumption data includes: workshop temperature and humidity, power consumption, steam consumption, compressed air pressure.
[0006] Further, the specific method of preprocessing is to clean, normalize and preliminarily format the collected raw data to form a standardized raw feature data set; Data cleaning includes: rejecting outliers and filling missing values; Preliminary format conversion includes: converting image data into a pixel matrix and converting time series data into a fixed length sequence.
[0007] Further, the step S3 specifically includes the following steps: S31, first-level feature extraction: for time series data, sliding window method + statistical analysis is used to extract basic time series features including mean, variance, peak value, valley value, trend slope, and cycle frequency; for structured data, one-hot encoding and cross statistics are used to generate structured basic features; for unstructured data, lightweight CNN is used to extract low-order visual features to output fixed-dimension feature vectors; S32, second-level feature mining based on first-level features: calculate the correlation coefficient of "device vibration peak value x main shaft temperature" and "product size error", and the mutual information value of "cutting torque fluctuation" and "appearance defect occurrence rate", generate key association features, construct the mapping relationship features of "production cycle length x device load rate" and "unit product power consumption", and the matching degree features of "order batch size" and "steam consumption peak value", use attention mechanism to weight all first-level basic features, highlight the features that have significant impact on the target, and generate fusion second-level feature vectors; S33, third-level feature mining based on second-level features: for time series association features in the second-level features, use LSTM / TCN model to extract trend prediction features such as "future 1 hour correlation coefficient change trend" and "potential abnormal inflection point probability" as time series prediction features, use GNN to abstract the factory as a "device-production-quality-energy consumption" association graph, take the second-level features as node attributes, extract bottleneck features of global resource allocation and collaborative optimization space features as global optimization features, use a classifier to pre-train fault association features in the second-level features to generate "fault type probability vector", and generate "quality level prediction vector" as high-order semantic features.
[0008] Further, the production scheduling optimization model adopts reinforcement learning, takes global optimization features in the three-level features as input, takes "minimum production cycle + highest device utilization" as reward function, and outputs optimal device allocation and process sorting instructions; The quality prediction and parameter optimization model takes "quality grade prediction vector" and "device-quality correlation high-order feature" in the three-level features as input, predicts quality risk through a CNN-LSTM model, and then outputs optimized production parameters through gradient descent reverse output; The device failure early warning model takes "failure type probability vector" and "time series prediction feature" in the three-level features as input, predicts failure 1-48 hours in advance through a Transformer model, and outputs failure grade and maintenance suggestions The energy consumption optimization model takes "production-energy consumption correlation high-order feature" and "global optimization feature" in the three-level features as input, adopts a genetic algorithm + linear regression hybrid model, and outputs dynamic energy consumption adjustment parameters.
[0009] Further, in step S5, multi-source data of the factory is collected in real time, preprocessed, processed into three-level features, input into the AI model to generate control instructions, and the control instructions are issued, such as adjusting processing parameters, updating MES scheduling, and optimizing energy consumption device parameters.
[0010] Further, in step S6, the result data after executing the control instructions is collected, including yield, failure rate, and energy consumption, and the effectiveness of the three-level features is evaluated. If the result does not meet the expectation, the feedback data is added to the data set, the three-level feature extraction algorithm is re-optimized, and the AI model parameters are updated using incremental learning.
[0011] According to another aspect of the present application, an AI-based digital factory intelligent control system is provided, which is used to implement the AI-based digital factory intelligent control method described above, and includes a data collection module, a data preprocessing module, a three-level feature processing module, a multi-task AI model module, a control execution module, and a feedback optimization module. The data collection module is used to collect device operation data, production process data, quality detection data, environment and energy consumption data of the factory, and transmit the raw data to the data preprocessing module. The data preprocessing module is used to clean, normalize, and preliminarily format the raw data, and output a standardized raw feature data set. The three-level feature processing module is used to extract features from the raw feature data set according to data types to obtain first-level features, generate second-level features based on the management relationship and weighting of the first-level features, obtain three-level features based on the extraction of high-order semantics and prediction features from the second-level features, and receive instructions from the feedback optimization module to optimize the feature extraction algorithm. Multi-task AI model module: used to train four types of models, production scheduling optimization, quality prediction and parameter optimization, equipment fault early warning, and energy consumption optimization, with three-level features as input, output control instructions to control execution module, and receive instructions from feedback optimization module to execute incremental learning update; Control execution module: used to receive and execute control instructions; Feedback optimization module: used to collect execution result data of control instructions, evaluate the effectiveness of three-level features, and issue instructions to three-level feature processing module to optimize feature extraction algorithm and multi-task AI module to perform incremental learning update.
[0012] Beneficial effects: three-level feature extraction mechanism breaks down data silos and mines high-order semantic information, first-level features realize dimensionality reduction and information condensation of raw data, second-level features capture cross-dimensional causal relationships, and third-level features are converted into decision-level knowledge. Four types of models, production scheduling, quality prediction, fault early warning, and energy consumption optimization, are integrated to realize multi-objective collaborative decision-making: for example, when the fault early warning model suggests "reducing device load", the production scheduling model will adjust the load of other devices to ensure the project schedule, and the energy consumption optimization model will match the adjusted load to optimize energy consumption parameters, avoiding the instruction conflict of traditional single-task models.
[0013] Through the feedback optimization module, real-time control effect data (yield, fault occurrence rate, energy consumption) is collected, the contribution of three-level features is evaluated (such as adjusting feature weights based on SHAP values), and incremental learning is used to update model parameters, without the need to retrain full data. When the factory experiences equipment aging, process upgrading, or order structure changes, the system can complete feature and model iteration within 24 hours, ensuring long-term stability of control precision. By predicting quality in advance, avoiding defective products, reducing downtime losses through fault early warning, dynamically optimizing energy consumption to reduce energy consumption, and reducing manual intervention costs. BRIEF DESCRIPTION OF DRAWINGS
[0014] Figure 1 is a method flow diagram. DETAILED DESCRIPTION
[0015] To make the technical solutions of the present application clearer, the following will further describe the present application in detail in combination with the drawings and specific embodiments. EMBODIMENT
[0016] 1. System deployment and data collection (corresponding to "data collection module") Hardware selection: Device operation data: CNC machine deployed PCB352C33 vibration sensor (sampling rate 1 kHz), OMRONE5CC infrared temperature sensor (accuracy ±0.1℃), LEMLA55-P current sensor, collecting spindle speed (500-10000 r / min), X / Y / Z axis vibration frequency (0-500 Hz), motor current (0-30 A), fault code (such as "E01 spindle overload"); Production process data: Production rhythm (120-180 s / piece), raw material ratio (cast iron powder: alloy: additive = 92:6:2), process progress (such as "cylinder drilling completion rate 75%") collected by Siemens S7-1200 PLC, order information (such as "high priority, batch 500 pieces, delivery period 72h") obtained through MES system; Quality detection data: Keyence IV2 vision camera (resolution 2048x1536) is used to detect appearance defects (scratches, depressions), and Hexagon Global Silver three-coordinate measuring instrument (accuracy ±0.001 mm) is used to collect cylinder hole size (φ80±0.02 mm); Environmental and energy consumption data: Sensirion SHT31 temperature and humidity sensor (15-35℃, 30%-70%RH), Schneider iEM3100 intelligent electric meter (0-100kW), ABB FST030 steam flowmeter (0-500kg / h).
[0017] Data transmission: Profinet IRT industrial Ethernet (cycle 1 ms) is used to transmit to Siemens IoT2050 edge gateway, and through MQTT protocol (QoS2) to upload to factory private cloud, vibration, current, etc. High frequency data sampling rate 1 kHz, temperature, rhythm, etc. Medium frequency data 1 Hz.
[0018] 2. Data preprocessing (corresponding to "data preprocessing module") Outlier rejection: vibration data uses 3σ principle (mean 12 Hz, standard deviation 3 Hz, data >21 Hz or <3 Hz is rejected), size error uses box plot method (data >80.03 mm or <79.97 mm outside 1.5 times of interquartile range is rejected); Missing value filling: Short-term missing (≤5s) uses linear interpolation (such as fitting and filling based on 10 vibration data points before and after), long-term missing (>5s) uses data migration of the same model 5# machine tool (such as when 2# machine tool temperature data is missing, use 5# machine tool temperature data at the same period to correct); Standardization and Format Conversion: Temperature (20-80℃) is normalized to [0, 1] by Min-Max normalization (formula: x_norm=(x-20) / (80-20)), order priority "high / medium / low" is one-hot encoded (high→[1,0,0]), appearance defect image is resized to 224x224 pixels and converted to grayscale image, output standardized original feature dataset.
[0019] 3. Three-level Feature Extraction (Corresponding to "Three-level Feature Processing Module") (1) First-level Feature Extraction Time Series Data (Vibration, Temperature): Sliding window size 10s, step 5s, extract 18-dimensional features such as vibration mean (12Hz), peak (18Hz), trend slope (0.5Hz / s), temperature mean (55℃), variance (2℃²); Structured Data (Order, Proportion): Order batch 500 pieces, raw material proportion 92:6:2, generate "batch × proportion coefficient" (500 × 0.92 = 460) and other 12-dimensional features; Unstructured Data (Appearance Image): MobileNetV2 extracts 256-dimensional low-order visual features (edge, texture).
[0020] (2) Second-level Feature Extraction Device-Quality Association: Calculate the Pearson correlation coefficient (0.82) of "vibration peak × temperature" (18 × 55 = 990) and dimensional error (80.01mm), and the mutual information value (0.75) of "cutting torque fluctuation" and appearance defect occurrence rate (2%); Production-Energy Consumption Association: Construct "production rhythm 150s × device load rate 80%" and unit power consumption (8.5kW・h / piece) mapping features (150 × 0.8 = 120, corresponding to power consumption 8.5); Attention Fusion: Weight the first-level features (vibration-related feature weight 0.35, environmental humidity weight 0.05), generate 14-dimensional second-level feature vector.
[0021] (3) Third-level Feature Extraction Time Series Prediction: LSTM model (hidden layer 32 dimensions) predicts future 1-hour trend (0.87, rising) and abnormal inflection point probability (0.12) for "correlation coefficient time series (0.82, 0.83, 0.85)"; Global Optimization: GNN abstracts "device-production-quality" association graph, extracts "3# machine tool load rate 90% leads to global delay risk 0.75" bottleneck features; Semantic features: SVM generates a failure probability vector (bearing wear 0.88, motor overload 0.10) and a quality grade vector (excellent 0.91, good 0.08, poor 0.01), outputting a 10-dimensional three-level feature.
[0022] 4. Multi-task AI model training and control execution (1) Model training Production scheduling optimization model: PPO algorithm (discount factor 0.95, learning rate 1e-4), input global bottleneck features, output "3# machine tool undertakes 60% task, 7# machine tool undertakes 40%" scheduling; Quality prediction model: CNN-LSTM (convolution kernel 3x3, LSTM hidden layer 64), input quality grade vector, predict poor rate 0.9%, output "main shaft speed from 850r / min to 820r / min" in reverse; Fault warning model: Transformer (encoder 3 layers), input failure probability vector, 24 hours in advance warning "3# machine tool bearing wear", suggest "replace bearing within 12 hours"; Energy optimization model: genetic algorithm (population 50, iteration 50 generations), input production-energy correlation features, output "air conditioner temperature 26℃, steam valve opening 75%";
[0023] (2) Control execution Edge end issues instructions to PLC: 3# machine tool speed adjustment to 820r / min, frequency converter frequency from 50Hz to 45Hz; MES system updates scheduling: 3# machine tool adds 300 orders, 7# machine tool adds 200; Energy consumption equipment response: air conditioner temperature from 24℃ to 26℃, steam valve opening from 80% to 75%.
[0024] 5. Closed-loop feedback and iteration Feedback data: 24 hours later, the cylinder size poor rate decreased from 3.5% to 0.8%, 3# machine tool bearing replacement failure did not occur, unit energy consumption decreased from 8.5kW・h / piece to 7.6kW・h / piece; Feature and model optimization: "environmental humidity" feature contribution (SHAP value) decreased from 0.1 to 0.04, adjust the first-level feature extraction weight; incremental learning updates quality prediction model parameters, next time prediction accuracy improves to 94%. Implementation effect
[0025] Product yield: from 96.5% to 99.2%, monthly reduction of 210 defective products, saving rework cost 180,000 yuan; Equipment failure: downtime from 8h per month to 3h, reducing production capacity loss 450,000 yuan; Energy cost: Unit energy consumption decreased by 10.6%, resulting in monthly electricity savings of 68,000 yuan. Example
[0026] 1. System Deployment and Data Acquisition Hardware selection: Equipment operating data: The reactor is equipped with a platinum resistance temperature sensor (0-300℃, accuracy ±0.5℃), a pressure transmitter (0-10MPa), and a stirring motor current sensor; Production process data: The raw material analyzer collects ethylene purity (99.5%-99.9%) and hydrogen / ethylene ratio (0.02-0.05), and the pelletizing machine collects pelletizing speed (500-800 r / min). Quality inspection data: Melt flow index (0.8-3.0g / 10min) of the product was collected by melt flow indexer (190℃ / 2.16kg), and particle size (2-5mm) was detected by particle size analyzer. Energy consumption data: Smart meters collect data on the heating power of the reactor (0-200kW) and the power consumption of the circulating water pump (0-50kW).
[0027] Data transmission: Transmission to Rockwell Edge Server is carried out using Modbus TCP protocol. Sampling rate: 1Hz for temperature and pressure, 5min / sampling for melt flow index.
[0028] 2. Data Preprocessing Outlier removal: Values with reaction temperatures exceeding 180-220℃ (process range) are directly removed. The melt index follows the 3σ principle (average 1.8g / 10min, data >3.0 or <0.6 are removed). Missing value imputation: For missing hydrogen ratio values (≤10 min), the average of adjacent time periods was used for imputation. Standardization: The reaction temperature (180-220℃) is normalized to [0,1], and the ethylene purity (99.5%-99.9%) is converted to a "purity coefficient" (e.g., 99.7%→0.7).
[0029] 3. Three-level feature extraction (1) Primary features Time-series data: 15-dimensional characteristics including mean reaction temperature (205℃), pressure variance (0.3MPa²), peak heating power (180kW); Structured data: Ethylene purity 0.7, hydrogen ratio 0.035, generating 10-dimensional features such as "purity × ratio" (0.7 × 0.035 = 0.0245).
[0030] (2) Secondary features Quality-parameter correlation: Calculate the correlation coefficient (0.86) of "reaction temperature x pressure" (205 x 5.5 = 1127.5) and melt index (1.9 g / 10 min); Energy consumption-load correlation: Calculate the mutual information value (0.81) of "heating power x pelletizing speed" (180 x 650 = 117000) and unit energy consumption (12 kW·h / kg); Attention fusion: Generate 12-dimensional secondary features (temperature-pressure correlation feature weight 0.4).
[0031] (3) Tertiary features Time series prediction: TCN model predicts the melt index trend in the next 2 hours (1.85→1.92); Global optimization: GNN extracts the collaborative feature of "1# reaction kettle heating power being too high causing the circulating water pump load to increase"; Semantic features: CNN-LSTM generates quality probability vector (melt index qualified 0.93, over-standard 0.07), outputs 9-dimensional tertiary features.
[0032] 4. Multi-task AI model and control execution Quality prediction model: Input quality probability vector, predict melt index over-standard risk 7%, output "reaction temperature from 205℃ to 203℃, hydrogen ratio from 0.035 to 0.037"; Energy consumption optimization model: Genetic algorithm outputs "heating power from 180kW to 172kW, circulating water pump frequency from 50Hz to 48Hz"; Control execution: DCS system receives instructions, adjusts reaction kettle temperature control valve and hydrogen feed valve in real time, and maintains pelletizer cutting speed at 650r / min. Implementation effect
[0033] Product quality: Melt index qualified rate increased from 92% to 98.5%, reduced unqualified products by 75 tons per month, saved losses of 450,000 yuan; Energy consumption cost: Unit product energy consumption decreased from 12kW·h / kg to 10.8kW·h / kg, saved electricity bill 120,000 yuan per month; Operation efficiency: Artificial parameter adjustment frequency decreased from 8 times per day to 2 times per day, operation and maintenance workload reduced by 75%.
[0034] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as limiting the scope of the present patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the present patent should be subject to the appended claims.
Claims
1. An AI-based digital plant intelligent control method, characterized by, The method specifically comprises the following steps: S1, multi-source data collection: collecting equipment operation, production process, quality detection, environment and energy consumption data; S2, multi-source data preprocessing: generating an original feature data set after preprocessing the collected multi-source data through cleaning and normalization; S3, three-level feature extraction: extracting features from the original feature data set according to data types to obtain first-level features, generating second-level features based on the management relationship of the first-level features, and obtaining third-level features based on the extraction of high-order semantics and prediction features from the second-level features; S4, multi-task AI model training: taking the third-level features as input, training four types of models including production scheduling optimization, quality prediction and parameter optimization, equipment fault early warning, and energy consumption optimization; S5, real-time monitoring and execution: collecting multi-source data in real time, inputting the preprocessed and three-level deep feature extracted data into the AI model to generate and issue control instructions; S6, closed-loop feedback and model iteration: collecting execution results, evaluating the effectiveness of the three-level features, optimizing the feature extraction algorithm, and incrementally updating the AI model parameters.
2. The AI-based digital plant intelligent control method according to claim 1, characterized in that: The multi-source data includes: equipment operation data, production process data, quality detection data, and environment and energy consumption data; The equipment data includes: equipment speed, current, temperature, vibration frequency, and fault code; The production process data includes: order quantity, production rhythm, raw material ratio, and process progress; The quality detection data includes: product size error, appearance defect, and performance parameter; The environment and energy consumption data includes: workshop temperature and humidity, power consumption, steam consumption, and compressed air pressure.
3. The AI-based digital plant intelligent control method according to claim 1, characterized in that: The specific method of preprocessing is to clean, normalize, and preliminarily format the collected raw data to form a standardized original feature data set; Data cleaning includes: removing outliers and filling missing values; Preliminary format conversion includes: converting image data into a pixel matrix and converting time series data into a fixed length sequence.
4. The AI-based digital plant intelligent control method according to claim 1, characterized in that: The step S3 specifically comprises the following steps: S31, first-level feature extraction: for time series data, sliding window method + statistical analysis is used to extract basic time series features including mean, variance, peak value, valley value, trend slope, and cycle frequency; for structured data, one-hot encoding and cross statistics are used to generate structured basic features; for unstructured data, lightweight CNN is used to extract low-order visual features to output a fixed-dimension feature vector; S32, mining second-level features based on first-level features: calculating the correlation coefficient of "equipment vibration peak value x main shaft temperature" and "product size error", and the mutual information value of "cutting torque fluctuation" and "appearance defect occurrence rate", generating key correlation features, constructing mapping relationship features of "production rhythm length x equipment load rate" and "unit product power consumption", and matching degree features of "order batch size" and "steam consumption peak value", using attention mechanism to weight all first-level basic features, highlighting features that significantly affect the target, and generating a fusion second-level feature vector; S33, based on the secondary feature mining three-level features: for the time sequence correlation features in the secondary features, the trend prediction features such as "the change trend of the correlation coefficient in the future 1 hour" and "the probability of potential abnormal inflection point" are extracted by using LSTM / TCN model, the factory is abstracted as "equipment-production-quality-energy consumption" correlation graph by using GNN, the secondary features are used as node attributes, and the bottleneck features of global resource allocation and collaborative optimization space features are extracted as global optimization features, the fault correlation features in the secondary features are pre-trained by using the classifier to generate "fault type probability vector", and the quality correlation features are generated as "quality level prediction vector" as high-order semantic features.
5. The AI-based digital plant intelligent control method according to claim 4, characterized in that: The production scheduling optimization model adopts reinforcement learning, takes the global optimization features in the three-level features as input, takes "minimum production cycle + highest equipment utilization rate" as reward function, and outputs optimal equipment allocation and process sequencing instructions; The quality prediction and parameter optimization model takes "quality level prediction vector" and "equipment-quality correlation high-order feature" in the three-level features as input, predicts quality risk through CNN-LSTM model, and reversely outputs optimized production parameters combined with gradient descent; The equipment fault early warning model takes "fault type probability vector" and "time sequence prediction feature" in the three-level features as input, predicts faults 1-48 hours in advance through the Transformer model, and outputs fault level and maintenance suggestions The energy consumption optimization model takes "production-energy consumption correlation high-order feature" and "global optimization feature" in the three-level features as input, adopts genetic algorithm + linear regression hybrid model, and outputs dynamic energy consumption adjustment parameters.
6. The AI-based digital plant intelligent control method according to claim 1, characterized in that: In the step S5, the real-time factory multi-source data is collected, preprocessed, and input into the AI model after three-level feature processing to generate control instructions and issue control instructions such as adjusting processing parameters, updating MES scheduling, and optimizing energy consumption equipment parameters.
7. The AI-based digital plant intelligent control method according to claim 1, characterized in that: In the step S6, the result data after collecting the control instructions is collected, including yield, fault occurrence rate, and energy consumption, while the effectiveness of the three-level features is evaluated, if the result does not meet the expectation, the feedback data is added to the data set, the three-level feature extraction algorithm is re-optimized, and the AI model parameters are updated by using incremental learning.
8. An AI-based digital plant intelligent control system, characterized in that: The system is used to realize the AI-based digital factory intelligent control method of any one of claims 1-7, comprising: a data acquisition module, a data preprocessing module, a three-level feature processing module, a multi-task AI model module, a control execution module, and a feedback optimization module; The data acquisition module is used to collect equipment operation data, production process data, quality detection data, environment and energy consumption data of the factory, and transmit the raw data to the data preprocessing module; The data preprocessing module is used to clean, normalize and preliminarily format the data, and output the standardized raw feature data set; The third-level feature processing module is configured to extract first-level features from the original feature dataset according to data types, generate second-level features based on the management relationship of the first-level features, obtain third-level features based on the second-level features and high-order semantics and prediction features, and receive instructions from the feedback optimization module to optimize the feature extraction algorithm. The multi-task AI model module is configured to take the third-level features as input, train four types of models including production scheduling optimization, quality prediction and parameter optimization, equipment fault early warning and energy consumption optimization, output control instructions to control the execution module, and receive instructions from the feedback optimization module to perform incremental learning and update. The control execution module is configured to receive and execute the control instructions. The feedback optimization module is configured to collect execution result data of the control instructions, evaluate the effectiveness of the third-level features, and issue instructions to the third-level feature processing module to optimize the feature extraction algorithm and to the multi-task AI model module to perform incremental learning and update.