Visual management and control method and system for intelligent production of pre-cooled concrete

Abstract

The invention discloses a visual management and control method and system for intelligent production of pre-cooled concrete. The method comprises the following steps: constructing a three-dimensional visual digital twinborn model through a camera, a sensor and a three-dimensional digital twinborn modeling method; decomposing a production structure of the three-dimensional visual digital twinborn model into local process modules; inputting field target production project data into the three-dimensional visual digital twinborn model to carry out real-time visual production simulation; performing first production control optimization on the field target production project through a neural network; performing secondary production management and control optimization on the field target production project through the three-dimensional visual digital twinborn model; a production field real-time scene is output through the three-dimensional visual digital twinborn model, and third-time production management and control optimization is carried out through a cosine similarity algorithm; through deep learning model prediction and monitoring, the neural network is introduced to control production personnel, intelligent management of the whole concrete production process is realized, the concrete quality is improved, and the safety risk is reduced.

Description

Technical Field

[0001] The present invention relates to the field of intelligent visual control technology, and in particular to a method and system for visual control of intelligent production of pre-cooled concrete. Background Art

[0002] With the development of the economy and society, the concrete industry continues to grow rapidly. To better achieve the goal of intelligent development, it is necessary to implement intelligent management and control of all aspects of concrete production. Currently, the concrete production process is basically equipped with an information management system. Through this system, seamlessly integrated with the computer control system of the concrete mixing plant, the scale system of the scale building, and the laboratory raw material and concrete product testing and inspection system, a series of data processing operations such as raw material testing, issuing production orders, uploading mix ratio notifications to the mixing plant control room, and concrete product quality control are realized through network automation and intelligentization. The entire process of weighing, conveying, mixing, and discharging in the concrete production process is automatically completed by the coordinated cooperation of computers, their electrical peripherals, and mechanical equipment. Therefore, the current concrete production has only achieved a certain level of automation, but cannot be adjusted and controlled by artificial intelligence according to actual production conditions, raw material quality, and performance requirements. The quality of the product is completely dependent on the experience of the operators.

[0003] Chinese patent document CN217688025U describes an online unmanned automatic sampling device for concrete. The technical application field of this patent is different from that of the present invention. It focuses on the automation of concrete sampling and does not cover the intelligent and visual management of the entire concrete production process. At the same time, it does not introduce intelligent management of sudden health conditions of on-site production personnel caused by some harsh production environments. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for visually controlling the intelligent production of pre-cooled concrete, so as to solve the above-mentioned problems existing in the prior art.

[0005] The specific application is as follows:

[0006] A method for visually controlling intelligent production of pre-cooled concrete, comprising the following steps:

[0007] S1. Construct a 3D visual digital twin model of the pre-cooling concrete production structure using cameras, sensors, and 3D digital twin modeling methods.

[0008] S2. Decompose the production structure of the 3D visual digital twin model into local process modules using a modular design method based on 3D digital twins. The local process modules include a raw material pre-cooling treatment scenario module, a proportioning and metering scenario module, and a mixing process control scenario module.

[0009] S3. Input the on-site target production project data into the 3D visual digital twin model for real-time visual production simulation;

[0010] S4. Analyze and predict the equipment factors, raw material factors, environmental factors, personnel factors, and local process modules of the on-site target production project using a neural network optimization algorithm within the three-dimensional visual digital twin model to generate a first prediction value for performing a first production control optimization. The first prediction value includes at least a production risk value.

[0011] S5. After the first production control optimization, start the pre-cooled concrete production of the on-site target production project;

[0012] S6. Using a 3D visualized digital twin model to collect and monitor real-time data on pre-cooled concrete production for the target on-site production project, and perform a secondary production control optimization, including using a neural network model to manage the health of equipment and production personnel.

[0013] S7. Output the real-time scene of the production site through the three-dimensional visualization digital twin model, calculate the similarity between the scene of the local process module under the real-time scene and the preset risk history scene of the local process module through the cosine similarity algorithm, and perform the third production control optimization based on the similarity result. The real-time scene includes several of the local process modules.

[0014] Furthermore, step S1 includes:

[0015] Preprocess real-time physical pre-cooling concrete equipment information and use the pre-processed data to simulate and model the pre-cooling concrete production structure, generating a 3D visual digital twin model. Preprocessing includes data cleaning, cluster analysis, iterative updates, and time series splicing.

[0016] The simulation modeling of the pre-cooling concrete production structure includes: using UG software to construct and simulate a model of the physical equipment based on the state data of the physical equipment during operation, combined with the material parameters and geometric parameters of the physical equipment, and visually displaying the state of the constructed physical equipment model in real time;

[0017] The three-dimensional visual digital twin model includes a database, a computing unit and a visualization unit;

[0018] The database is used to store data output by the three-dimensional visualization digital twin model, the physical equipment model constructed by the three-dimensional visualization digital twin model, model parameters transmitted by the AI ​​algorithm model library, and historical operating status data of the physical equipment on the physical pre-cooling concrete equipment;

[0019] The computing unit is configured to read the real-time physical production information of the physical pre-cooling concrete equipment from a database, call the model parameters in the AI ​​algorithm model library from the database, perform abnormality detection and analysis on the real-time production information of the physical equipment on the physical pre-cooling concrete equipment using the AI ​​algorithm model, and output the analysis and processing results to the visualization unit to display the real-time production information of the physical pre-cooling concrete equipment;

[0020] The visualization unit is used for visual display of the three-dimensional visualization digital twin model, and real-time status marking of the production information of the physical pre-cooling concrete equipment. The status marking is divided into a first status, a second status, and a third status. The first status mark is green, indicating that the equipment is operating normally; the second status mark is blue, indicating that the equipment will be operating abnormally; the third status mark is red, indicating that the equipment is operating abnormally.

[0021] Furthermore, step S2 includes:

[0022] S21, production process decomposition;

[0023] By creating component element associations through the object-relationship model, the entire pre-cooled concrete intelligent production process is decomposed according to standardized process requirements;

[0024] S22, local process module design;

[0025] The parametric modeling unit of UG software is used to define at least the equipment shape, equipment size, equipment connection method, and equipment material requirement parameters of each process module, and then different local process modules are constructed through the parametric modeling unit. The local process modules include a raw material pre-cooling treatment scenario module, a proportioning and metering scenario module, and a mixing process control scenario module;

[0026] S23, local process module identification;

[0027] RFID technology is used to encode each local process module with a unique identification code for quick identification and installation at the pre-cooled concrete intelligent production site;

[0028] S24, production process simulation optimization;

[0029] The intelligent production of pre-cooled concrete is simulated and optimized through the three-dimensional design model Blender to determine the optimal production sequence and method; the three-dimensional design model Blender simulates the local process module design process through a 4D simulation method.

[0030] Furthermore, step S4 includes:

[0031] Based on a neural network optimization algorithm, a pre-cooled concrete intelligent production time prediction function is used to predict the impact of equipment factors, raw material factors, environmental factors, and personnel matching factors on the production time of the on-site target production project. The pre-cooled concrete intelligent production time prediction function collects and analyzes a large amount of historical project production data through a neural network algorithm, and predicts the impact of different equipment factors, raw material factors, environmental factors, and personnel matching factors on the production time through a Bayesian regression method. The expression of the pre-cooled concrete intelligent production time prediction function is:

[0032]

[0033] Among them, T1 represents the predicted intelligent production time of pre-cooled concrete; p is the equipment factor, which is used to indicate the degree of equipment loss; S is the raw material factor, which is used to indicate the adaptability of raw materials; θ is the environmental factor, which is used to indicate the comprehensive value of the production weather, the slope of the production site, and the distance from the production site to the concrete use site; K is the personnel matching factor, which is used to indicate the suitability of the production project personnel matching; d is the weight influence coefficient;

[0034] Based on a neural network optimization algorithm, a pre-cooled concrete intelligent production risk prediction function is used to predict the impact of each local process module on the production risk of the on-site target production project. The pre-cooled concrete intelligent production risk prediction function collects and analyzes a large amount of historical project production risk data through a neural network algorithm, and predicts the impact of each local process module on production risk through a decision tree regression method. The expression of the pre-cooled concrete intelligent production risk prediction function is:

[0035]

[0036] Among them, M represents the total production risk value, which is used to evaluate the safety risk level in the production process; M i represents the production risk value of the i-th local process, y i is the probability risk coefficient of the i-th local process module, which is used to indicate the probability of occurrence of the local process; x i is the consequence risk coefficient of the i-th local process module, which is used to indicate the risk level of the impact after the local process occurs; z i is the frequency risk coefficient of the i-th local process module, which is used to indicate the number of repetitions of the local process.

[0037] Furthermore, in step S6, real-time data collection and monitoring of pre-cooled concrete production of the on-site target production project is performed through the three-dimensional visual digital twin model to perform a second production control optimization, including:

[0038] S61. Capture facial images, walking posture images, and voice data of production personnel through cameras, and acquire basic equipment information and monitoring information through sensors. The facial images include historically acquired images and real-time acquired images; the basic information includes equipment type, model, manufacturer, and date of manufacture; the monitoring information includes mechanical vibration values ​​of the mixer and conveyor belt acquired through vibration sensors; temperature values ​​of the motor, hydraulic system, and bearings monitored through temperature sensors; pressure fluctuation values ​​of the hydraulic system and pneumatic system acquired through pressure sensors; and noise values ​​of gear wear and component collisions acquired through noise sensors.

[0039] S62. Develop a production personnel health management model by training a first neural network model using historical and real-time data on facial images, walking posture images, and voice data. Develop an equipment health management model by training a second neural network model using historical data on basic equipment information and monitoring information. Both the production personnel health management model and the equipment health management model are embedded in a three-dimensional visual digital twin model.

[0040] S63. Use the trained equipment health management model to monitor the health of the equipment of the target production project on site in real time, and use the output data for the second production control optimization. Use the trained production personnel health management model to monitor the health of the production personnel of the target production project on site in real time, and use the output data for the second production control optimization.

[0041] Furthermore, the specific process of training the first neural network model in step S62 is as follows:

[0042] Preprocess the facial images, walking posture images, and voice data to obtain the data corresponding to the preprocessed production personnel;

[0043] Performing feature extraction on the preprocessed facial image, walking posture image, and voice data through feature engineering to generate facial image features, walking posture image features, and voice data features. Part of the extracted facial image features, walking posture image features, and voice data features are used for a training set, and part is used for a test set. The feature extraction includes shallow feature extraction and deep feature extraction. The deep feature extraction is based on the shallow feature extraction, and the accuracy of the deep feature extraction is greater than that based on the shallow feature extraction.

[0044] The first neural network model is trained using the data of the training set, and the training is stopped when the accuracy of the first neural network model in determining that the health status of the production personnel has changed reaches a threshold;

[0045] The first neural network model is specifically:

[0046] Suppose there are three inputs: face image features F a , walking posture image features F b and speech data features F c , define the first neural network model of the three layers respectively, let g(F a ) represents the output of the face image feature level, g(F b ) represents the output of the walking posture image feature level, g(F c ) represents the output of the speech data feature level, then:

[0047] DD=Connect(g(F a ), g(F b ), g(F c ));

[0048] P(Y=i2|F a , F b , F c )=Sigmoid((f(DD) i3 );

[0049] Among them, DD represents the three-layer neural network g(F a ), g(F b ), g(F c )’s output is connected together in a new dimension to obtain a vector; Connect means connecting multiple features in a new dimension to generate a longer vector; P(Y=i2|F a , F b , F c ) represents the given face image feature F a , walking posture image features F b and speech data features F c In the case of , the probability that the sample belongs to the i2-th category feature; f(DD) represents the neural network layer that further processes the concatenated vector DD; (f(DD) i3 ) represents the value of the i3th element in the vector output by the first neural network model f; Sigmoid represents the activation function.

[0050] Furthermore, in step S63, the health of production personnel of the on-site target production project is monitored in real time by using the trained production personnel health management model, and the output data is used for the second production management optimization, including:

[0051] S631. Set the threshold for changes in the health status of production personnel to Q. Initially configure the feature extraction method of the production personnel health management and control model to shallow feature extraction. If the output result of the production personnel health management and control model is greater than the threshold Q, the system performs a second production management and control optimization, which at least includes notifying the production personnel to rest or providing emergency medical assistance.

[0052] S632. Set the threshold for changes in the health status of production personnel to Q1. If the output result of the production personnel health management model is greater than the threshold Q1 and less than Q, the production personnel health management model continues to track the production personnel and extracts features through deep feature extraction. Set the time for continuous tracking of production personnel to T2. During T2, the camera transmits the image to the production personnel health management model every 10ms. If the average value of the output result of the production personnel health management model within T2 is less than the threshold Q1, the continuous tracking of the production personnel is canceled, and the camera is changed to transmit the image to the production personnel health management model every 100ms. The feature extraction method of the production personnel health management model is changed to shallow feature extraction, where Q1 is less than Q.

[0053] Furthermore, step S7 includes:

[0054] S71. If a first threshold is preset, and if the production risk value of the i-th local process module in the real-time scene of the production site is greater than the preset first threshold, output the scene corresponding to the i-th local process module through the three-dimensional visual digital twin model;

[0055] S72. Analyze the comparison between the scenario of the i-th local process module in the real-time scenario and the preset risk historical scenario of the i-th local process module using a cosine similarity algorithm to generate a comparison value, where the preset risk historical scenario is represented as a historical scenario containing a risk warning;

[0056] S73. Preset a second threshold value. If the comparison value meets the second threshold value, generate a fault alarm and execute the third production control optimization.

[0057] A pre-cooled concrete intelligent production visualization management and control system includes at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor performs any method described in a pre-cooled concrete intelligent production visualization management and control method.

[0058] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects:

[0059] The embodiment of the present invention provides a method for constructing a 3D visual digital twin model of a pre-cooled concrete production structure by using cameras, sensors and a 3D digital twin modeling method, thereby realizing visual management and control design of the entire process; decomposing the production structure of the 3D visual digital twin model into local process modules by using a module design method based on the 3D digital twin; inputting the on-site target production project data into the 3D visual digital twin model for real-time visual production simulation; analyzing and predicting the equipment factors, raw material factors, environmental factors, personnel matching factors and local process modules of the on-site target production project by using the optimization algorithm of the neural network in the 3D visual digital twin model, generating a first prediction value, and performing the first production management and control optimization; after the first production management and control optimization After the initialization, the pre-cooling concrete production of the on-site target production project is started; the real-time data of the pre-cooling concrete production of the on-site target production project is collected and monitored through the three-dimensional visualization digital twin model, and the second production control optimization is carried out; the real-time scene of the production site is output through the three-dimensional visualization digital twin model, and the scene of the local process module under the real-time scene is analyzed by the cosine similarity algorithm to compare the preset risk history scene of the local process module, and the third production control optimization is carried out according to the comparison results; based on digital twin technology, the present invention uses deep learning model prediction and monitoring, introduces neural networks to control sudden health problems of production personnel, realizes intelligent and visual management of the entire concrete production process, improves concrete quality, and reduces safety risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 The present invention provides a flow chart of a method for visually controlling the intelligent production of pre-cooled concrete. DETAILED DESCRIPTION

[0061] The present invention will be described in detail below with reference to the accompanying drawings.

[0062] Example 1

[0063] The embodiment of the present invention provides a method for visual control of intelligent production of pre-cooled concrete. Figure 1 , the method comprising:

[0064] S1. Construct a 3D visual digital twin model of the pre-cooling concrete production structure using cameras, sensors, and 3D digital twin modeling methods.

[0065] S2. Decompose the production structure of the 3D visual digital twin model into local process modules using a modular design method based on 3D digital twins. The local process modules include a raw material pre-cooling treatment scenario module, a proportioning and metering scenario module, and a mixing process control scenario module.

[0066] S3. Input the on-site target production project data into the 3D visual digital twin model for real-time visual production simulation;

[0067] S4. Analyze and predict the equipment factors, raw material factors, environmental factors, personnel factors, and local process modules of the on-site target production project using a neural network optimization algorithm within the three-dimensional visual digital twin model to generate a first prediction value for performing a first production control optimization. The first prediction value includes at least a production risk value.

[0068] S5. After the first production control optimization, start the pre-cooled concrete production of the on-site target production project;

[0069] S6. Using a 3D visualized digital twin model to collect and monitor real-time data on pre-cooled concrete production for the target on-site production project, and perform a secondary production control optimization, including using a neural network model to manage the health of equipment and production personnel.

[0070] S7. Output the real-time scene of the production site through the three-dimensional visualization digital twin model, calculate the similarity between the scene of the local process module under the real-time scene and the preset risk history scene of the local process module through the cosine similarity algorithm, and perform the third production control optimization based on the similarity result. The real-time scene includes several of the local process modules.

[0071] Specifically, this embodiment uses digital twin technology to design a full-process visual management and control system for on-site target production projects. Through deep learning model prediction and monitoring, neural networks are introduced to control sudden health problems of production personnel and the health of equipment, thereby realizing intelligent and visual management of the entire concrete production process, improving concrete quality, and reducing safety risks.

[0072] In the above embodiment, specifically, step S1 includes:

[0073] Preprocess real-time physical pre-cooling concrete equipment information and use the pre-processed data to simulate and model the pre-cooling concrete production structure, generating a 3D visual digital twin model. Preprocessing includes data cleaning, cluster analysis, iterative updates, and time series splicing.

[0074] Preferably, the specific steps of cluster analysis in preprocessing (for preprocessing data of precooling concrete equipment)

[0075] 1. Data preparation stage

[0076] Input data:

[0077] Real-time sensor data (temperature, vibration, current, pressure);

[0078] Equipment status log (start and stop events, fault codes);

[0079] Environmental parameters (ambient temperature and humidity, cooling water flow);

[0080] Feature Engineering:

[0081] Time domain characteristics: mean, variance, crest factor;

[0082] Frequency domain characteristics: FFT extracts the dominant frequency amplitude (such as the characteristic frequency of bearing fault);

[0083] Time series statistics: Sliding window calculation of trend slope (such as temperature change rate ΔT / Δt);

[0084] 2. Clustering algorithm selection and optimization

[0085] Algorithm selection:

[0086] DBSCAN: Applicable to device data with high noise (automatically removes outliers);

[0087] K-means++: for steady-state operating data (the number of clusters K needs to be preset and determined by the elbow rule);

[0088] GMM (Gaussian Mixture Model): handles multimodal distributions (such as temperature clusters at different cooling stages);

[0089] Industrial adaptation improvements:

[0090] Dynamic weight adjustment: assign higher feature weights to key sensors (such as kettle temperature);

[0091] Incremental clustering: performs online updates on newly arriving data streams (e.g., sampling every minute) to avoid full recalculation.

[0092] 3. Core clustering process

[0093] Step 1: Noise filtering and data cleaning

[0094] Use Isolation Forest to remove obvious outliers (such as spikes caused by sensor failure);

[0095] Example: Filter invalid sampling points > 10g in vibration data (normal range 0.1-2g);

[0096] Step 2: Multi-dimensional feature reduction

[0097] PCA (Principal Component Analysis): compresses high-dimensional features (such as 20-dimensional sensor data) into 3-dimensional principal components, retaining 95% of the variance;

[0098] t-SNE: Visualize clustering effects (used in the debugging phase);

[0099] Step 3: Clustering Execution and Label Generation

[0100] DBSCAN parameter settings:

[0101] Neighborhood radius ε: determined by the k-distance graph (e.g., ε = 0.5);

[0102] Minimum number of samples MinPts: Adjust according to the equipment type (set to 15 for mixer and 10 for conveyor belt);

[0103] Output cluster labels:

[0104] Cluster 0: normal stirring state (stable temperature, small current fluctuation);

[0105] Cluster 1: potential bearing wear (high-frequency vibration components are prominent);

[0106] Cluster 2: Insufficient cooling (temperature drop rate is lower than the threshold);

[0107] Step 4: Clustering result verification

[0108] Silhouette Score: evaluates cluster tightness (target > 0.6);

[0109] Manual annotation review: Compare with historical fault records and correct misclassified clusters;

[0110] 4. Clustering output and digital twin mapping

[0111] 3D Visualization Association:

[0112] Each cluster corresponds to a different colored area in the digital twin model (e.g., red = high-risk cluster);

[0113] Dynamically annotate cluster features (e.g., “Cluster 1: Vibration peak 500Hz, bearing inspection recommended”);

[0114] Iterative update mechanism:

[0115] Re-cluster every 24 hours and update cluster centers and boundaries;

[0116] Model retraining is triggered when new data is collected and the silhouette coefficient is less than 0.5.

[0117] The simulation modeling of the pre-cooling concrete production structure includes: using UG software to construct and simulate a model of the physical equipment based on the state data of the physical equipment during operation, combined with the material parameters and geometric parameters of the physical equipment, and visually displaying the state of the constructed physical equipment model in real time;

[0118] The three-dimensional visual digital twin model includes a database, a computing unit and a visualization unit;

[0119] The database is used to store data output by the three-dimensional visualization digital twin model, the physical equipment model constructed by the three-dimensional visualization digital twin model, model parameters transmitted by the AI ​​algorithm model library, and historical operating status data of the physical equipment on the physical pre-cooling concrete equipment;

[0120] The computing unit is configured to read the real-time physical production information of the physical pre-cooling concrete equipment from a database, call the model parameters in the AI ​​algorithm model library from the database, perform abnormality detection and analysis on the real-time production information of the physical equipment on the physical pre-cooling concrete equipment using the AI ​​algorithm model, and output the analysis and processing results to the visualization unit to display the real-time production information of the physical pre-cooling concrete equipment;

[0121] The visualization unit is used for visual display of the three-dimensional visualization digital twin model, and real-time status marking of the production information of the physical pre-cooling concrete equipment. The status marking is divided into a first status, a second status, and a third status. The first status mark is green, indicating that the equipment is operating normally; the second status mark is blue, indicating that the equipment will be operating abnormally; the third status mark is red, indicating that the equipment is operating abnormally.

[0122] Specifically, the visualization unit also includes:

[0123] Graphical presentation:

[0124] Topology: This diagram uses a force-directed graph to show the relationship between devices and fault nodes, with the root cause node highlighted in red.

[0125] Timeline: Overlay the time series of multiple alarm events to identify concurrent or chained faults;

[0126] 3D overlay: Annotate the physical location of the fault in the three-dimensional visual digital twin model (e.g., click on a blender to display a bearing wear hotspot);

[0127] Interactive features:

[0128] Drill down analysis: click on any node to view detailed data (such as vibration spectrum, maintenance records);

[0129] Hypothesis deduction: Manually modify parameters (such as "if the speed is reduced") to simulate whether the fault is eliminated;

[0130] Report generation: Automatically output root cause analysis report (including treatment suggestions and spare parts list).

[0131] In the above embodiment, specifically, step S2 includes:

[0132] S21, production process decomposition;

[0133] By creating component element associations through the object-relationship model, the entire pre-cooled concrete intelligent production process is decomposed according to standardized process requirements;

[0134] S22, local process module design;

[0135] The parametric modeling unit of UG software is used to define at least the equipment shape, equipment size, equipment connection method, and equipment material requirement parameters of each process module, and then different local process modules are constructed through the parametric modeling unit. The local process modules include a raw material pre-cooling treatment scenario module, a proportioning and metering scenario module, and a mixing process control scenario module;

[0136] S23, local process module identification;

[0137] RFID technology is used to encode each local process module with a unique identification code for quick identification and installation at the pre-cooled concrete intelligent production site;

[0138] S24, production process simulation optimization;

[0139] The intelligent production of pre-cooled concrete is simulated and optimized through the three-dimensional design model Blender to determine the optimal production sequence and method; the three-dimensional design model Blender simulates the local process module design process through a 4D simulation method.

[0140] In the above embodiment, specifically, step S4 includes:

[0141] Based on a neural network optimization algorithm, a pre-cooled concrete intelligent production time prediction function is used to predict the impact of equipment factors, raw material factors, environmental factors, and personnel matching factors on the production time of the on-site target production project. The pre-cooled concrete intelligent production time prediction function collects and analyzes a large amount of historical project production data through a neural network algorithm, and predicts the impact of different equipment factors, raw material factors, environmental factors, and personnel matching factors on the production time through a Bayesian regression method. The expression of the pre-cooled concrete intelligent production time prediction function is:

[0142]

[0143] Among them, T1 represents the predicted intelligent production time of pre-cooled concrete; p is the equipment factor, which is used to indicate the degree of equipment loss; S is the raw material factor, which is used to indicate the adaptability of raw materials; θ is the environmental factor, which is used to indicate the comprehensive value of the production weather, the slope of the production site, and the distance from the production site to the concrete use site; K is the personnel matching factor, which is used to indicate the suitability of the production project personnel matching; d is the weight influence coefficient;

[0144] Based on a neural network optimization algorithm, a pre-cooled concrete intelligent production risk prediction function is used to predict the impact of each local process module on the production risk of the on-site target production project. The pre-cooled concrete intelligent production risk prediction function collects and analyzes a large amount of historical project production risk data through a neural network algorithm, and predicts the impact of each local process module on production risk through a decision tree regression method. The expression of the pre-cooled concrete intelligent production risk prediction function is:

[0145]

[0146] Among them, M represents the total production risk value, which is used to evaluate the safety risk level in the production process; M i represents the production risk value of the i-th local process, y i is the probability risk coefficient of the i-th local process module, which is used to indicate the probability of occurrence of the local process; x i is the consequence risk coefficient of the i-th local process module, which is used to indicate the risk level of the impact after the local process occurs; z i is the frequency risk coefficient of the i-th local process module, which is used to indicate the number of repetitions of the local process.

[0147] Specifically, the neural network optimization algorithm includes but is not limited to BP neural network algorithm, decision tree neural network algorithm, etc.

[0148] It should be noted that the specific steps of performing the first production control optimization in step S4 also include:

[0149] a. Equipment factor optimization

[0150] Optimization goal: Improve equipment operating efficiency and reduce failure risks;

[0151] Implementation steps:

[0152] Health status prediction:

[0153] Analyze vibration and current time series data through LSTM neural networks to predict the remaining useful life (RUL) of key equipment (such as mixer motors);

[0154] Output the first prediction value: Equipment Health Index (EHI)∈[0,1], when EHI<0.6, trigger maintenance warning;

[0155] Dynamic parameter adjustment:

[0156] Optimize the combination of mixer speed and duration based on reinforcement learning (PPO algorithm) to balance energy consumption and concrete homogeneity;

[0157] Example: When the concrete temperature is predicted to rise too quickly, the speed is automatically reduced by 10% and the mixing time is extended by 15%;

[0158] b. Raw material factor optimization

[0159] Optimization goal: Ensure accurate material ratio and reduce cooling loss;

[0160] Implementation steps:

[0161] Moisture content compensation:

[0162] The aggregate moisture content is detected in real time by a microwave moisture sensor, and the neural network dynamically corrects the water consumption (ΔW calculation formula: ΔW = 0.8 × ΔMC + 0.2 × T_env, ΔMC is the change in moisture content, and T_env is the ambient temperature);

[0163] Pre-cooling strategy adjustment:

[0164] Based on the cluster analysis of infrared thermal imaging sensor data, the uneven areas of aggregate pre-cooling are identified and the air outlet angle and wind speed of the air cooler are adjusted;

[0165] Output the first predicted value: the deviation between the target temperature and the actual temperature of the aggregate is ≤±0.5℃;

[0166] c. Environmental factors optimization

[0167] Optimization goal: offset the impact of environmental fluctuations on temperature control;

[0168] Implementation steps:

[0169] Temperature and humidity compensation model:

[0170] Input environmental sensor data and use GNN (graph neural network) to predict the concrete outlet temperature change curve T(t);

[0171] When the predicted T(t) exceeds the threshold, the amount of borneol added is automatically increased (2kg / m2 of borneol is added for every 1°C increase). 3 );

[0172] Anti-wind control:

[0173] Combining weather station data with digital twin wind farm simulation, we optimized the opening and closing strategies for transport vehicle insulation covers to reduce cooling loss.

[0174] d. Optimization of personnel matching factors

[0175] Optimization goal: Improve the efficiency of human-machine collaboration

[0176] Implementation steps:

[0177] AR Operation Guide:

[0178] Key step prompts (e.g., "Current water-cement ratio is 0.38, it is recommended to reduce water addition by 5L") are superimposed on the operator through Hololens glasses;

[0179] Team effectiveness evaluation:

[0180] Based on computer vision-based behavioral analysis (such as operating rhythm and tool access time), the neural network outputs personnel efficiency scores and recommends the optimal team combination;

[0181] e. Collaborative optimization of local process modules

[0182] Optimization goal: global optimality instead of local optimality

[0183] Implementation steps:

[0184] Inter-module coupling analysis:

[0185] Mining association rules between proportion measurement and mixing process through knowledge graph (e.g., “sand ratio increased by 1% → mixing torque needs to be increased by 3N·m”);

[0186] Dynamic resource scheduling:

[0187] Predict the production capacity bottlenecks of each module (such as when the mixer utilization rate is >90%), and dispatch spare equipment or adjust the production rhythm in advance.

[0188] Furthermore, in step S6, real-time data collection and monitoring of pre-cooled concrete production of the on-site target production project is performed through the three-dimensional visual digital twin model to perform a second production control optimization, including:

[0189] S61. Capture facial images, walking posture images, and voice data of production personnel through cameras, and acquire basic equipment information and monitoring information through sensors. The facial images include historically acquired images and real-time acquired images; the basic information includes equipment type, model, manufacturer, and date of manufacture; the monitoring information includes mechanical vibration values ​​of the mixer and conveyor belt acquired through vibration sensors; temperature values ​​of the motor, hydraulic system, and bearings monitored through temperature sensors; pressure fluctuation values ​​of the hydraulic system and pneumatic system acquired through pressure sensors; and noise values ​​of gear wear and component collisions acquired through noise sensors.

[0190] S62. Develop a production personnel health management model by training a first neural network model using historical and real-time data on facial images, walking posture images, and voice data. Develop an equipment health management model by training a second neural network model using historical data on basic equipment information and monitoring information. Both the production personnel health management model and the equipment health management model are embedded in a three-dimensional visual digital twin model.

[0191] S63. Use the trained equipment health management model to monitor the health of the equipment of the target production project on site in real time, and use the output data for the second production control optimization. Use the trained production personnel health management model to monitor the health of the production personnel of the target production project on site in real time, and use the output data for the second production control optimization.

[0192] In the above embodiment, specifically, the specific process of training the first neural network model in step S62 is:

[0193] Preprocess the facial images, walking posture images, and voice data to obtain the data corresponding to the preprocessed production personnel;

[0194] Performing feature extraction on the preprocessed facial image, walking posture image, and voice data through feature engineering to generate facial image features, walking posture image features, and voice data features. Part of the extracted facial image features, walking posture image features, and voice data features are used for a training set, and part is used for a test set. The feature extraction includes shallow feature extraction and deep feature extraction. The deep feature extraction is based on the shallow feature extraction, and the accuracy of the deep feature extraction is greater than that based on the shallow feature extraction.

[0195] The first neural network model is trained using the data of the training set, and the training is stopped when the accuracy of the first neural network model in determining that the health status of the production personnel has changed reaches a threshold;

[0196] The first neural network model is specifically:

[0197] Suppose there are three inputs: face image features F a , walking posture image features F b and speech data features F c , define the first neural network model of the three layers respectively, let g(F a ) represents the output of the face image feature level, g(F b ) represents the output of the walking posture image feature level, g(F c ) represents the output of the speech data feature level, then:

[0198] DD=Connect(g(F a), g(F b ), g(F c ));

[0199] P(Y=i2|F a , F b , F c )=Sigmoid((f(DD) i3 );

[0200] Among them, DD represents the three-layer neural network g(F a ), g(F b ), g(F c )’s output is connected together in a new dimension to obtain a vector; Connect means connecting multiple features in a new dimension to generate a longer vector; P(Y=i2|F a , F b , F c ) represents the given face image feature F a , walking posture image features F b and speech data features F c In the case of , the probability that the sample belongs to the i2-th category feature; f(DD) represents the neural network layer that further processes the concatenated vector DD; (f(DD) i3 ) represents the value of the i3th element in the vector output by the first neural network model f; Sigmoid represents the activation function.

[0201] It should be noted that the facial image features include facial complexion features, facial lip color features and fatigue features;

[0202] The facial complexion features include red, yellow, black, white, pale, dark red and reddish yellow;

[0203] The facial lip color characteristics include facial lip white, facial lip dark red, facial lip light, facial lip red, facial lip bluish purple, facial lip purple and facial lip purple black;

[0204] The fatigue characteristics include eyelid closure frequency and pupil diameter changes;

[0205] The walking posture image features include stride length and stride frequency;

[0206] The speech data features include acoustic features, language features and semantic health condition keyword features.

[0207] In the above embodiment, specifically, in step S63, the health of production personnel of the on-site target production project is monitored in real time by using the trained production personnel health management and control model, and the output data is used for the second production management and control optimization, including:

[0208] S631. Set the threshold for changes in the health status of production personnel to Q. Initially configure the feature extraction method of the production personnel health management and control model to shallow feature extraction. If the output result of the production personnel health management and control model is greater than the threshold Q, the system performs a second production management and control optimization, which at least includes notifying the production personnel to rest or providing emergency medical assistance.

[0209] S632. Set the threshold for changes in the health status of production personnel to Q1. If the output result of the production personnel health management model is greater than the threshold Q1 and less than Q, the production personnel health management model continues to track the production personnel and extracts features through deep feature extraction. Set the time for continuous tracking of production personnel to T2. During T2, the camera transmits the image to the production personnel health management model every 10ms. If the average value of the output result of the production personnel health management model within T2 is less than the threshold Q1, the continuous tracking of the production personnel is canceled, and the camera is changed to transmit the image to the production personnel health management model every 100ms. The feature extraction method of the production personnel health management model is changed to shallow feature extraction, where Q1 is less than Q.

[0210] It should be noted that the specific steps of performing the second production control optimization in step S6 also include:

[0211] a. Equipment health management and optimization

[0212] a.1 Real-time fault diagnosis and predictive maintenance

[0213] Anomaly Detection:

[0214] The equipment health management model analyzes equipment sensor data (vibration, temperature, current, etc.) in real time, and uses a 2D-CNN+LSTM hybrid model to detect transient anomalies (e.g., using vibration sensors to obtain mechanical vibration values ​​of mixers and conveyor belts; using temperature sensors to monitor the temperature values ​​of motors, hydraulic systems, and bearings; using pressure sensors to obtain pressure fluctuation values ​​of hydraulic systems and pneumatic systems; and using noise sensors to obtain noise values ​​due to gear wear and component collisions).

[0215] Output indicator: Failure probability value (FPV∈[0,1]), triggering an alarm when FPV>0.7;

[0216] Remaining Life Prediction (RUL):

[0217] Based on a deep survival analysis model (in this example, the DeepSurv neural network model), combined with historical equipment maintenance records, the remaining service life of key components (such as the mixer reduction gearbox) is predicted;

[0218] Optimized actions: Automatically generate maintenance work orders and push them to the maintenance terminal, and simultaneously adjust production schedules;

[0219] a.2 Dynamic Energy Efficiency Optimization

[0220] Energy consumption-performance balance:

[0221] Input real-time power consumption data and concrete quality indicators (slump, strength) to optimize equipment operating parameters through multi-objective reinforcement learning (MO-PPO);

[0222] Example: Under the premise of ensuring strength, reducing the mixer speed by 5% can save energy consumption by 8%;

[0223] Load balancing scheduling:

[0224] Dynamically assign tasks based on equipment health status (e.g., prioritize mixers with an EHI > 0.8) to extend overall equipment life;

[0225] b. Optimization of production personnel health management

[0226] Calculated by the Health Risk Index (HRI):

[0227] The production personnel health management model (using a model such as the LightGBM ensemble algorithm) takes as input facial image features, walking posture image features, and voice data features, and outputs HRI, where HRI∈[0,1]:

[0228] If HRI ≥ 0.8: Emergency (e.g., heat stroke, heart discomfort), triggering audible and visual alarms, automatically calling the factory doctor, and immediately locking related equipment (e.g., emergency stop of the mixer);

[0229] If 0.6≤HRI<0.8: High-risk state (immediate intervention required), the AR glasses will send a rest instruction and automatically dispatch backup personnel to take over;

[0230] If HRI<0.6: normal monitoring.

[0231] In the above embodiment, specifically, step S7 includes:

[0232] S71. If a first threshold is preset, and if the production risk value of the i-th local process module in the real-time scene of the production site is greater than the preset first threshold, output the scene corresponding to the i-th local process module through the three-dimensional visual digital twin model;

[0233] It should be noted that if the production risk value of the i-th local process module in the real-time scene of the production site is less than or equal to the preset first threshold, the three-dimensional visualization digital twin model does not need to output the corresponding scene, which is beneficial to saving output resources and improving the three-dimensional visualization digital twin model's ability to flexibly output risky scenes for key monitoring.

[0234] S72. Analyze the comparison between the scenario of the i-th local process module in the real-time scenario and the preset risk historical scenario of the i-th local process module using a cosine similarity algorithm to generate a comparison value, where the preset risk historical scenario is represented as a historical scenario containing a risk warning;

[0235] S73. Preset a second threshold value. If the comparison value meets the second threshold value, generate a fault alarm and execute the third production control optimization.

[0236] In the above embodiment, specifically, the specific implementation process of step S72 is:

[0237] Vectorize the scenarios of local process modules and their preset risk history scenarios through vector engineering in the 3D visual digital twin model.

[0238] The cosine similarity algorithm is used to match and search all preset risk historical scenario data stored in the database to find the preset risk historical scenario with the highest similarity. The Word2Vec algorithm is used to extract the solution keywords corresponding to the preset risk historical scenario with the highest similarity. The whale algorithm is used to perform a web crawler search for the solution keywords to obtain the final solution for the third production control optimization.

[0239] Cosine similarity algorithm:

[0240]

[0241] The cosine similarity algorithm is used to calculate the similarity between the scenario of the ith local process module and the preset risk history scenario of the ith local process module. The larger the cosθ value, the more similar it is, and vice versa. i Represents the set of feature vectors corresponding to the scene of the i-th local process module, y i Represents the set of feature vectors corresponding to the preset risk history scenarios of the i-th local process module.

[0242] It should be noted that the specific steps of the third production control optimization in step S7 also include:

[0243] The third round of production control optimization was carried out through the visualization unit and root cause analysis dashboard of the 3D visualization digital twin model;

[0244] The root cause analysis dashboard is embedded in the three-dimensional visual digital twin model. The root cause analysis dashboard automatically executes the third production control optimization task in step S73 by constructing a multi-dimensional knowledge graph of equipment, process, and environment. If the system triggers the execution of the third production control optimization task, the root cause analysis dashboard associates real-time fault alarms with historical fault data, graphically displaying the root cause of the fault and the impact path, assisting operation and maintenance personnel in quickly locating the problem. At the same time, combined with the final solution obtained by the whale algorithm, the fault problem can be quickly resolved.

[0245] The multidimensional knowledge graph includes knowledge graph construction and graph modeling;

[0246] Knowledge graph construction:

[0247] Data source:

[0248] Static knowledge base: equipment topology relationship (such as mixer → conveyor belt → aggregate silo), process parameter dependency (such as the relationship between water-cement ratio and strength), historical fault case library;

[0249] Dynamic data stream: real-time sensor alarms (temperature exceeding the limit, vibration abnormality), production logs (operation records, maintenance records);

[0250] Graph Modeling:

[0251] Entity type: equipment, sensor, process parameter, environmental variable, failure mode (such as "bearing wear" and "motor overload");

[0252] Weight allocation: Based on the probability of fault occurrence and expert rules, the strength of the association between entities is quantified (e.g., "bearing wear → vibration abnormality" has a weight of 0.9);

[0253] Fault propagation path generation:

[0254] Dynamic Inference Engine:

[0255] Input: Real-time alarm events (e.g., “conveyor belt speed drops suddenly”);

[0256] Reasoning logic:

[0257] Subgraph matching: extracting entities and relationships directly connected to alarm events from the knowledge graph;

[0258] Probability propagation: Calculate root cause probabilities based on Bayesian networks (e.g., "aggregate blockage" probability 70% vs. "inverter failure" probability 30%);

[0259] Path tracing: trace back the causal chain to the most likely root cause and simulate the impact range forward;

[0260] The introduction of the root cause analysis dashboard achieves the technical innovation of this embodiment, which is:

[0261] Multi-source heterogeneous data fusion: Unify equipment topology, process parameters, and environmental data into a knowledge graph, breaking through the information silos of traditional alarm systems;

[0262] Dynamic probabilistic reasoning: Combines real-time data with historical statistics to dynamically adjust root cause weights and adapt to changes in the production environment.

[0263] Explainable visualization: By mapping interactive graphs with physical models, the "black box" threshold of AI algorithms is lowered.

[0264] Example 2

[0265] A pre-cooled concrete intelligent production visualization management and control system includes at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor performs any method described in a pre-cooled concrete intelligent production visualization management and control method.

[0266] It should be understood that the above embodiments are one or more embodiments of the present invention, and there are many other embodiments and variations thereof based on the present invention; the variations and modifications made by ordinary technicians in this industry through the present invention without making groundbreaking innovations all fall within the scope of protection of the present invention.

Claims

1. A method for visual control of intelligent production of pre-cooled concrete, characterized in that: The method comprises the following steps: S1. Construct a 3D visual digital twin model of the pre-cooling concrete production structure using cameras, sensors, and 3D digital twin modeling methods. S2. Decompose the production structure of the 3D visual digital twin model into local process modules using a modular design method based on 3D digital twins. The local process modules include a raw material pre-cooling treatment scenario module, a proportioning and metering scenario module, and a mixing process control scenario module. S3. Input the on-site target production project data into the 3D visual digital twin model for real-time visual production simulation; S4. Analyze and predict the equipment factors, raw material factors, environmental factors, personnel factors, and local process modules of the on-site target production project using a neural network optimization algorithm within the three-dimensional visual digital twin model to generate a first prediction value for performing a first production control optimization. The first prediction value includes at least a production risk value. S5. After the first production control optimization, start the pre-cooled concrete production of the on-site target production project; S6. Using a 3D visualized digital twin model to collect and monitor real-time data on pre-cooled concrete production for the target on-site production project, and perform a secondary production control optimization, including using a neural network model to manage the health of equipment and production personnel. S7. Output the real-time scene of the production site through the three-dimensional visualization digital twin model, calculate the similarity between the scene of the local process module under the real-time scene and the preset risk history scene of the local process module through the cosine similarity algorithm, and perform the third production control optimization based on the similarity result. The real-time scene includes several of the local process modules.

2. A method for visual control of intelligent production of pre-cooled concrete according to claim 1, characterized in that: Step S1 includes: Preprocess real-time physical pre-cooling concrete equipment information and use the pre-processed data to simulate and model the pre-cooling concrete production structure, generating a 3D visual digital twin model. Preprocessing includes data cleaning, cluster analysis, iterative updates, and time series splicing. The simulation modeling of the pre-cooling concrete production structure includes: using UG software to construct and simulate a model of the physical equipment based on the state data of the physical equipment during operation, combined with the material parameters and geometric parameters of the physical equipment, and visually displaying the state of the constructed physical equipment model in real time; The three-dimensional visual digital twin model includes a database, a computing unit and a visualization unit; The database is used to store data output by the three-dimensional visualization digital twin model, the physical equipment model constructed by the three-dimensional visualization digital twin model, model parameters transmitted by the AI ​​algorithm model library, and historical operating status data of the physical equipment on the physical pre-cooling concrete equipment; The computing unit is configured to read the real-time physical production information of the physical pre-cooling concrete equipment from a database, call the model parameters in the AI ​​algorithm model library from the database, perform abnormality detection and analysis on the real-time production information of the physical equipment on the physical pre-cooling concrete equipment using the AI ​​algorithm model, and output the analysis and processing results to the visualization unit to display the real-time production information of the physical pre-cooling concrete equipment; The visualization unit is used for visual display of the three-dimensional visualization digital twin model, and real-time status marking of the production information of the physical pre-cooling concrete equipment. The status marking is divided into a first status, a second status, and a third status. The first status mark is green, indicating that the equipment is operating normally; the second status mark is blue, indicating that the equipment will be operating abnormally; the third status mark is red, indicating that the equipment is operating abnormally.

3. The method for visual control of intelligent production of pre-cooled concrete according to claim 1, characterized in that: Step S2 includes: S21, production process decomposition; By creating component element associations through the object-relationship model, the entire pre-cooled concrete intelligent production process is decomposed according to standardized process requirements; S22, local process module design; The parametric modeling unit of UG software is used to define at least the equipment shape, equipment size, equipment connection method, and equipment material requirement parameters of each process module, and then different local process modules are constructed through the parametric modeling unit. The local process modules include a raw material pre-cooling treatment scenario module, a proportioning and metering scenario module, and a mixing process control scenario module; S23, local process module identification; RFID technology is used to encode each local process module with a unique identification code for quick identification and installation at the pre-cooled concrete intelligent production site; S24, production process simulation optimization; The intelligent production of pre-cooled concrete is simulated and optimized through the three-dimensional design model Blender to determine the optimal production sequence and method; the three-dimensional design model Blender simulates the local process module design process through a 4D simulation method.

4. The method for visual control of intelligent production of pre-cooled concrete according to claim 1, characterized in that: Step S4 includes: Based on a neural network optimization algorithm, a pre-cooled concrete intelligent production time prediction function is used to predict the impact of equipment factors, raw material factors, environmental factors, and personnel matching factors on the production time of the on-site target production project. The pre-cooled concrete intelligent production time prediction function collects and analyzes a large amount of historical project production data through a neural network algorithm, and predicts the impact of different equipment factors, raw material factors, environmental factors, and personnel matching factors on the production time through a Bayesian regression method. The expression of the pre-cooled concrete intelligent production time prediction function is: Among them, T1 represents the predicted intelligent production time of pre-cooled concrete; p is the equipment factor, which is used to indicate the degree of equipment loss; S is the raw material factor, which is used to indicate the adaptability of raw materials; θ is the environmental factor, which is used to indicate the comprehensive value of the production weather, the slope of the production site, and the distance from the production site to the concrete use site; K is the personnel matching factor, which is used to indicate the suitability of the production project personnel matching; d is the weight influence coefficient; Based on a neural network optimization algorithm, a pre-cooled concrete intelligent production risk prediction function is used to predict the impact of each local process module on the production risk of the on-site target production project. The pre-cooled concrete intelligent production risk prediction function collects and analyzes a large amount of historical project production risk data through a neural network algorithm, and predicts the impact of each local process module on production risk through a decision tree regression method. The expression of the pre-cooled concrete intelligent production risk prediction function is: Among them, M represents the total production risk value, which is used to evaluate the safety risk level in the production process; M i represents the production risk value of the i-th local process, y i is the probability risk coefficient of the i-th local process module, which is used to indicate the probability of occurrence of the local process; x i is the consequence risk coefficient of the i-th local process module, which is used to indicate the risk level of the impact after the local process occurs; z i is the frequency risk coefficient of the i-th local process module, which is used to indicate the number of repetitions of the local process.

5. The method for visual control of intelligent production of pre-cooled concrete according to claim 1, characterized in that: In step S6, real-time data collection and monitoring of pre-cooled concrete production of the on-site target production project is performed through a three-dimensional visual digital twin model, and a second production control optimization is performed, including: S61. Capture facial images, walking posture images, and voice data of production personnel through cameras, and acquire basic equipment information and monitoring information through sensors. The facial images include historically acquired images and real-time acquired images; the basic information includes equipment type, model, manufacturer, and date of manufacture; the monitoring information includes mechanical vibration values ​​of the mixer and conveyor belt acquired through vibration sensors; temperature values ​​of the motor, hydraulic system, and bearings monitored through temperature sensors; pressure fluctuation values ​​of the hydraulic system and pneumatic system acquired through pressure sensors; and noise values ​​of gear wear and component collisions acquired through noise sensors. S62. Develop a production personnel health management model by training a first neural network model using historical and real-time data on facial images, walking posture images, and voice data. Develop an equipment health management model by training a second neural network model using historical data on basic equipment information and monitoring information. Both the production personnel health management model and the equipment health management model are embedded in a three-dimensional visual digital twin model. S63. Use the trained equipment health management model to monitor the health of the equipment of the target production project on site in real time, and use the output data for the second production control optimization. Use the trained production personnel health management model to monitor the health of the production personnel of the target production project on site in real time, and use the output data for the second production control optimization.

6. A method for visual control of intelligent production of pre-cooled concrete according to claim 5, characterized in that: The specific process of the first neural network model training in step S62 is: Preprocess the facial images, walking posture images, and voice data to obtain the data corresponding to the preprocessed production personnel; Performing feature extraction on the preprocessed facial image, walking posture image, and voice data through feature engineering to generate facial image features, walking posture image features, and voice data features. Part of the extracted facial image features, walking posture image features, and voice data features are used for a training set, and part is used for a test set. The feature extraction includes shallow feature extraction and deep feature extraction. The deep feature extraction is based on the shallow feature extraction, and the accuracy of the deep feature extraction is greater than that based on the shallow feature extraction. The first neural network model is trained using the data of the training set, and the training is stopped when the accuracy of the first neural network model in determining that the health status of the production personnel has changed reaches a threshold; The first neural network model is specifically: Suppose there are three inputs: face image features F a , walking posture image features F b and speech data features F c , define the first neural network model of the three layers respectively, let g(F a ) represents the output of the face image feature level, g(F b ) represents the output of the walking posture image feature level, g(F c ) represents the output of the speech data feature level, then: DD=Connect(g(F a ),g(F b ),g(F c )); P(Y=i2|F a ,F b ,F c )=Sigmoid((f(DD) i3 ); Among them, DD represents the three-layer neural network g(F a ), g(F b ), g(F c )’s output is connected together in a new dimension to obtain a vector; Connect means connecting multiple features in a new dimension to generate a longer vector; P(Y=i2|F a , F b , F c ) represents the given face image feature F a , walking posture image features F b and speech data features F c In the case of , the probability that the sample belongs to the i2-th category feature; f(DD) represents the neural network layer that further processes the concatenated vector DD; (f(DD) i3 ) represents the value of the i3th element in the vector output by the first neural network model f; Sigmoid represents the activation function.

7. A method for visual control of intelligent production of pre-cooled concrete according to claim 6, characterized in that: In step S63, the health of production personnel of the on-site target production project is monitored in real time through the trained production personnel health management model, and the output data is used for the second production management optimization, including: S631. Set the threshold for changes in the health status of production personnel to Q. Initially configure the feature extraction method of the production personnel health management and control model to shallow feature extraction. If the output result of the production personnel health management and control model is greater than the threshold Q, the system performs a second production management and control optimization, which at least includes notifying the production personnel to rest or providing emergency medical assistance. S632. Set the threshold for changes in the health status of production personnel to Q1. If the output result of the production personnel health management model is greater than the threshold Q1 and less than Q, the production personnel health management model continues to track the production personnel and extracts features through deep feature extraction. Set the time for continuous tracking of production personnel to T2. During T2, the camera transmits the image to the production personnel health management model every 10ms. If the average value of the output result of the production personnel health management model within T2 is less than the threshold Q1, the continuous tracking of the production personnel is canceled, and the camera is changed to transmit the image to the production personnel health management model every 100ms. The feature extraction method of the production personnel health management model is changed to shallow feature extraction, where Q1 is less than Q.

8. The method for visual control of intelligent production of pre-cooled concrete according to claim 1, characterized in that: Step S7 includes: S71. If a first threshold is preset, and if the production risk value of the i-th local process module in the real-time scene of the production site is greater than the preset first threshold, output the scene corresponding to the i-th local process module through the three-dimensional visual digital twin model; S72. Analyze the comparison between the scenario of the i-th local process module in the real-time scenario and the preset risk historical scenario of the i-th local process module using a cosine similarity algorithm to generate a comparison value, where the preset risk historical scenario is represented as a historical scenario containing a risk warning; S73. Preset a second threshold value. If the comparison value meets the second threshold value, generate a fault alarm and execute the third production control optimization.

9. A pre-cooled concrete intelligent production visualization management and control system, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 8.

Images