An industrial internet production line dynamic optimization control method based on digital twinning
By constructing a synchronous mapping twin and a dynamic optimization model, dynamically allocating data fusion weights, and combining deep reinforcement learning and adaptive fuzzy logic, the problem of key information being submerged in multi-source heterogeneous data fusion is solved, thereby improving the accuracy and efficiency of dynamic optimization control of the production line.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SICHUAN YEXIN ENERGY INVESTMENT IND INTERNET CO LTD
- Filing Date
- 2025-10-25
- Publication Date
- 2026-06-19
Smart Images
Figure CN122239602A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial internet technology, specifically to a dynamic optimization control method for industrial internet production lines based on digital twins. Background Technology
[0002] In the intelligent management and control of industrial internet production lines, digital twin technology constructs a virtual mirror of the physical production line to achieve real-time mapping and dynamic optimization of the production line's operating status. Multi-source heterogeneous data fusion is the core link supporting the accurate modeling of digital twins. By integrating various types of data such as equipment sensors, environmental monitoring, and material management, it provides basic information for the production line's status perception and control decisions.
[0003] In existing technologies, multi-source heterogeneous data fusion often employs traditional machine learning methods (such as principal component analysis and Kalman filtering) or basic deep learning models (such as convolutional neural networks and recurrent neural networks). By using fixed weights or preset rules, features are extracted and fused from data from different sources. The fusion results are then used to construct a digital twin to reflect the overall operating status of the production line and provide data support for subsequent production scheduling, equipment control, and other tasks.
[0004] In existing technologies, during the fusion of multi-source heterogeneous data, the data fusion weights are often fixed due to the lack of dynamic correlation with the production line control strategy. This results in key data that significantly affects control decisions (such as core equipment parameters and quality inspection indicators) not being extracted first, and key information is easily overwhelmed by massive amounts of redundant data. This makes it difficult for the digital twin to accurately capture the core state of the production line, affecting the effectiveness of subsequent optimization control. In view of this, we propose a dynamic optimization control method for industrial internet production lines based on digital twins. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a dynamic optimization control method for industrial internet production lines based on digital twins. This method solves the problem that in existing technologies, data fusion weights are often fixed, resulting in the inability to prioritize the extraction of key data that significantly impacts control decisions.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a dynamic optimization control method for industrial internet production lines based on digital twins, comprising the following steps;
[0007] S1: Construct a synchronous mapping twin. Through real-time acquisition and fusion of multi-source heterogeneous data, generate a digital twin that maps to the physical production line and keep the two updated synchronously in real time.
[0008] S2: Build a dynamic optimization model. Based on the digital twin constructed in S1, construct a control model that integrates deep reinforcement learning and adaptive fuzzy logic. Generate multiple sets of control strategy candidate schemes based on the preset optimization target.
[0009] S3: Embedded predictive maintenance feedback, relying on the data accumulated in the digital twin in S1 to predict the device status, and feed the prediction results back to the dynamic optimization model in S2 to correct the candidate solution.
[0010] S4: Perform virtual-real interaction verification. Use the digital twin in S1 to perform virtual simulation of the candidate solutions modified in S3, select the optimal strategy and send it to the physical production line. At the same time, the physical production line data is continuously fed back to the digital twin in S1 to form a closed loop.
[0011] S5: Construct a twin interaction platform, build a visual interactive interface based on the digital twin of S1, and support manual intervention and adjustment of candidate solutions generated by S2, prediction results of S3 and verification process of S4.
[0012] Preferably, in the construction of the synchronous mapping twin in S1, the fusion of multi-source heterogeneous data adopts an improved attention mechanism deep learning model. This model introduces the control strategy influence factor output by the dynamic optimization model in S2 as the attention weight adjustment parameter. The data fusion weight is dynamically allocated according to the degree of influence of the control strategy on each link of the production line, thereby improving the feature extraction accuracy of the data.
[0013] Preferably, in the S1 construction of the synchronous mapping twin, the multi-source heterogeneous data includes vibration values, temperature readings, and pressure parameters collected by the equipment sensors; temperature and humidity indicators and dust concentration data obtained by the environmental monitoring device; material type, quantity, and conveying speed information recorded by the material management system; and appearance parameters, size data, and defect identifiers generated by the product testing equipment.
[0014] Preferably, in the S2 construction of the dynamic optimization model, the deep reinforcement learning adopts the improved DDPG algorithm that integrates the data fusion features of S1. The data features extracted in S1 are used as the input dimension of the state space, so that the agent pays priority to the data features that have a significant impact on the control effect during the policy learning process. At the same time, the real-time deviation of data fusion is included as a penalty term in the reward function to enhance the robustness of the model to data synchronization delay.
[0015] Preferably, in the S2 dynamic optimization model, the adaptive fuzzy logic adapts to the time-varying characteristics of the production line by adjusting the membership function parameters online. The preset optimization targets include the production cycle time compliance rate, unit product energy consumption value, and product sampling pass rate. The working conditions simulated by the digital twin include material supply fluctuations, equipment load changes, and environmental parameter drifts.
[0016] Preferably, S3 is embedded in the predictive maintenance feedback. The predictive maintenance module adopts a bidirectional long short-term memory network. The input data includes the equipment's historical operating parameters for the past 3 months and real-time status data for the past 72 hours. The output results include equipment fault type prediction, fault occurrence time interval, and remaining effective running time. The feedback method is to transmit the correction coefficient matrix to the dynamic optimization model.
[0017] Preferably, in the S3 embedded predictive maintenance feedback, the correction process includes recalculating the operating parameters of the equipment involved in the predicted failure in the candidate schemes, rearranging the timing of the coordinated actions of the associated equipment, and readjusting the material allocation scheme. After correction, the number of candidate schemes remains consistent with the initial number generated.
[0018] Preferably, in the virtual-real interaction verification performed in S4, the virtual simulation verification includes single-factor variable testing, multi-condition combination testing, and extreme state simulation testing. The verification indicators include strategy execution time, total resource consumption, and number of products meeting quality standards. The selection rule is to determine the scheme with the highest comprehensive score as the optimal strategy.
[0019] Preferably, in the S4 virtual-real interaction verification, the optimal strategy is distributed via an edge computing node, the transmission protocol is MQTT, and the physical production line feedback data includes the actual equipment parameters, actual production indicators, and actual quality test results after the strategy is executed. The closed-loop update cycle is once per hour.
[0020] Preferably, in the construction of the twin interaction platform in S5, the twin interaction platform adopts three-dimensional real-time rendering technology, supports gesture control and voice command input, and the manual intervention includes adjusting the optimization target weight in S2, modifying the prediction threshold in S3, and pausing the strategy issuance process in S4.
[0021] This invention provides a dynamic optimization control method for industrial internet production lines based on digital twins. It has the following beneficial effects:
[0022] 1. This invention utilizes an improved attention mechanism deep learning model in S1 to introduce dynamic allocation of data fusion weights based on control strategy influence factors, thereby enhancing the accuracy of key data feature extraction. After feature mapping and weighted fusion of multi-source heterogeneous data, the resulting global feature tensor accurately reflects the core state of the production line, solving the problem of key information being obscured in traditional data fusion. This provides high-quality data support for subsequent optimization control and improves the mapping accuracy between the digital twin and the physical production line.
[0023] 2. This invention utilizes an improved DDPG algorithm that integrates features from S1 into S2, incorporating the integrated features into the state space. Combined with a reward function that includes a synchronization delay penalty, this allows the agent to prioritize key features, enhancing robustness to data latency. The dynamic reward mechanism, linked with experience replay, improves the targeting and adaptability of control strategy generation, effectively optimizing targets such as production efficiency, energy consumption, and yield rate.
[0024] 3. This invention utilizes the bidirectional linkage of algorithms S1 and S2. S1 optimizes attention weights using the influence factors of the S2 control strategy, while S2 optimizes strategy learning based on the fusion features and delayed data from S1, forming a closed loop of "data features - strategy generation - feedback optimization". This linkage strengthens the pervasive role of key features in data fusion and strategy learning, improving the overall efficiency and synergy of dynamic optimization control on the production line. Attached Figure Description
[0025] Figure 1 This diagram illustrates a dynamic optimization control method for industrial internet production lines based on digital twins. Detailed Implementation
[0026] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] Example:
[0028] Please see the appendix Figure 1 This invention provides a method for dynamic optimization control of industrial internet production lines based on digital twins, comprising the following steps;
[0029] S1: Construct a synchronous mapping twin. Through real-time acquisition and fusion of multi-source heterogeneous data, generate a digital twin that maps to the physical production line and keep the two updated synchronously in real time.
[0030] In the synchronous mapping twin construction in S1, the fusion of multi-source heterogeneous data adopts an improved attention mechanism deep learning model. This model introduces the control strategy influence factor output by the dynamically optimized model in S2 as the attention weight adjustment parameter. By dynamically allocating data fusion weights according to the degree of influence of the control strategy on each link of the production line, the feature extraction accuracy of the data is improved. The following algorithms are included:
[0031] The first step is multi-source heterogeneous data preprocessing and tensor construction.
[0032] First, the collected n types of multi-source heterogeneous data (such as device sensor data, environmental data, etc.) are standardized to eliminate dimensional differences, and then converted into a tensor form suitable for model input, denoted as:
[0033]
[0034] in:
[0035] Let i represent the feature tensor of the i-th type of data source, with dimension n. ;
[0036] t is the time step;
[0037] Let be the feature dimension of the i-th data type.
[0038] For example, when i=1 represents temperature sensor data, It could be a tensor containing 100 time steps, with 8 temperature values per time step.
[0039] The second step is feature mapping in the feature extraction layer.
[0040] For each type of data, a corresponding feature extraction network is selected (e.g., convolutional neural networks handle image data, recurrent neural networks handle time-series data), and the feature extraction function is used. Will Mapped to a high-dimensional feature tensor ,Right now:
[0041]
[0042] in:
[0043] The dimension is k is a uniform feature dimension;
[0044] is the feature extraction function for the i-th type of data (such as a combination of convolutional and pooling layers).
[0045] For example, temperature time series data , An LSTM network can be used to output a feature tensor that includes the temperature change trend. .
[0046] The third step is to adapt the dimensions of the factors influencing the control strategy.
[0047] Obtain m control strategy influencing factors from the dynamic optimization model of S2.
[0048]
[0049] in:
[0050] This represents the influence coefficient of the j-th control strategy (such as the equipment speed adjustment strategy) on the production line (the value ranges from 0 to 1, and the larger the value, the more significant the influence).
[0051] Transform it into a vector that matches the dimension of the feature tensor through a linear transformation. The formula is:
[0052]
[0053] in:
[0054] It is a dimension of The trainable weight matrix (k is the feature dimension, m is the number of control policies);
[0055] It is a bias vector of dimension k;
[0056] This is a k-dimensional adaptation vector after transformation, ensuring compatibility with the feature tensor. Perform element-level operations;
[0057] Step 4: Calculation of improved attention weights
[0058] Based on feature tensor Influence factors of the adapted control strategy Calculate the attention score for each feature tensor. The formula is:
[0059]
[0060] in:
[0061] It is a dimension of The feature weight matrix;
[0062] It is a dimension of Attention weight matrix;
[0063] tanh is an activation function used to enhance nonlinear expressive power.
[0064] The scores were then normalized to attention weights using the softmax function. ,Right now:
[0065]
[0066] in:
[0067] Let the weights of the i-th type of feature tensor during fusion satisfy the following condition: And suffer Influence
[0068] When a certain type of data is highly correlated with the control strategies of high-impact factors... It will automatically increase in size;
[0069] Step 5: Weighted Feature Fusion and Output
[0070] The feature tensors are weighted and summed according to their attention weights to obtain the fused global feature tensor. The formula is:
[0071]
[0072] in:
[0073] The dimension is It is a fusion feature that contains all the key information of the data.
[0074] For example, if the equipment vibration data corresponding Larger (due to the influence of associated equipment maintenance strategies) (higher), then exist The proportion of this is larger, highlighting its reference value for control strategies;
[0075] Step 6: Joint optimization of model parameters
[0076] To minimize the loss function: To achieve this, all trainable parameters are updated via backpropagation. (etc.). The reconstruction loss is calculated using the following formula:
[0077]
[0078] in:
[0079] Let be the value of the fused feature in the q-th dimension at time step p;
[0080] This is the true label for that value (annotated by the best historical data);
[0081] To optimize the loss for attention, the formula is:
[0082] Feature importance calculated from attention weights ( for (L2 norm);
[0083] It is the expected importance of the j-th type of feature, determined based on expert experience;
[0084] It is a balance coefficient (with a value of 0.1-1) used to adjust the proportion of the two losses.
[0085] This loss function enables the model to improve the accuracy of feature reconstruction while ensuring that the influence of attention weights and control strategies remains consistent, thus achieving strong correlation optimization between the S1 and S2 algorithms.
[0086] In the synchronous mapping twin constructed in S1, the multi-source heterogeneous data includes vibration values, temperature readings, and pressure parameters collected by equipment sensors, temperature and humidity indicators and dust concentration data obtained by environmental monitoring devices, material type, quantity, and conveying speed information recorded by the material management system, and appearance parameters, size data, and defect identification generated by product testing equipment.
[0087] S2: Build a dynamic optimization model. Based on the digital twin constructed in S1, construct a control model that integrates deep reinforcement learning and adaptive fuzzy logic. Generate multiple sets of control strategy candidate schemes based on the preset optimization target.
[0088] In the S2-based dynamic optimization model, deep reinforcement learning employs an improved DDPG algorithm that integrates data fusion features from S1. The data features extracted from S1 are used as the input dimension of the state space, enabling the agent to prioritize data features that significantly impact control effectiveness during policy learning. Simultaneously, real-time deviations in data fusion are incorporated as a penalty term into the reward function, enhancing the model's robustness to data synchronization delays. This includes the following algorithms:
[0089] The first step is to construct an extended state space that incorporates S1 features.
[0090] The state vector is defined as ,in It is the fused feature sheet (dimension k) output by S1, containing key features (such as equipment vibration, temperature, etc.) enhanced by the S1 attention mechanism; It is the attention weight calculated in S1 (with dimension n, where n is the number of data sources), which directly reflects the degree of influence of each data on the control strategy; It is the synchronization delay time (in milliseconds) between the digital twin and the physical entity in S1, provided by the S1 real-time monitoring module.
[0091] The second step is to initialize the dual-network structure with S1 weight adaptation.
[0092] An S1 weight adaptation module is added to the input layer of the Actor policy network to adjust the attention weights. Transform into an adjustment matrix that matches the dimension of the state space. The conversion formula is:
[0093]
[0094] in:
[0095] It is a trainable weight matrix of dimension (k,n) (k is the feature dimension of Z, and n is the number of attention weights);
[0096] It is a bias vector of dimension k;
[0097] This indicates that a vector is transformed into a diagonal matrix;
[0098] Let be a kxk diagonal matrix, whose diagonal elements are... Positive correlation ensures that high-weight features in S1 are prioritized in the network;
[0099] The inputs to both the Actor network and the Critic network are processed through... Weighted state vector:
[0100]
[0101] in This indicates vector concatenation, making and Participating in calculation as an independent feature
[0102] The third step is to improve the forward propagation process of the Actor policy network.
[0103] The control policy output by the policy network is:
[0104]
[0105] in:
[0106] It is the set of Actor network parameters (including) )
[0107] It is the weight matrix of the first hidden layer (dimension 1). , (Number of neurons in the hidden layer)
[0108] For the corresponding bias;
[0109] It is the output layer weight matrix (dimension 1). , To control the dimensions of motion, such as the amount of equipment speed adjustment).
[0110] For output bias;
[0111] For activation functions;
[0112] Map the output to interval;
[0113] The function limits the result to the actual control range. Internal (e.g., speed adjustment from -50 to +50 r / min).
[0114] because It incorporates of The matrix and strategy generation process will automatically generate a matrix. High-weight features of the label are skewed.
[0115] Step 4, design including The dynamic reward function for delayed penalties. The formula for the reward function is as follows:
[0116]
[0117] in:
[0118] It is a production performance reward ;
[0119] To achieve the production efficiency target rate;
[0120] Energy consumption exceeding standards rate;
[0121] For the pass rate;
[0122] The weights are 1;
[0123] It is the delay penalty coefficient (value ranges from 0.02 to 0.2).
[0124] yes The synchronization delay time; the greater the delay, the heavier the penalty.
[0125] It is the attention alignment penalty coefficient (values range from 0.1 to 1).
[0126] It is the attention weight vector predicted by the Critic network (output through an additional branch);
[0127] The penalty term, representing the L1 norm of both, prompts the model to learn with... Feature attention patterns with consistent attention weights.
[0128] The fifth step is to build an experience replay mechanism that links with S1.
[0129] The sample format stored in the experience replay pool is as follows: ,in It is the attention weight of S1 at time t. It is the synchronization delay at time t, ensuring that the sample contains complete S1 dynamic features.
[0130] Sampling from the playback pool uses priority sampling based on S1 features, with sampling probabilities as follows:
[0131]
[0132] in:
[0133] It is an adjustment coefficient (with a value of 0.5-2);
[0134] It is the maximum value of the S1 attention weight in the sample at time t. This mechanism makes samples with high-weight features preferentially sampled, which enhances the learning effect of key features.
[0135] Step 6: Perform dual-network collaborative update and S1 feedback iteration.
[0136] The target value of the Critic network is calculated as follows:
[0137]
[0138] in:
[0139] It is a discount factor (values 0.9-0.99).
[0140] and It is a target Critic and Actor network;
[0141] Fusion features at time t+1 including S1 output
[0142] The loss function of the Critic network is:
[0143]
[0144] in:
[0145] N is the batch size;
[0146] These are the parameters of the main Critic network;
[0147] The gradient update for the Actor network is:
[0148]
[0149] in:
[0150] It is the mean of the S1 attention weights at time t, used to adjust the gradient magnitude so that samples dominated by high-weight features contribute more to the policy update.
[0151] After each iteration, the control strategy influence factor output by the Actor network is fed back to S1 as a reference for updating the attention weight of S1, forming a closed-loop optimization between S1 and S2.
[0152] In the S2 dynamic optimization model, adaptive fuzzy logic adapts to the time-varying characteristics of the production line by adjusting the membership function parameters online. The preset optimization objectives include production cycle time achievement rate, unit product energy consumption value, and product sampling pass rate. The working conditions simulated by the digital twin include material supply fluctuations, equipment load changes, and environmental parameter drift.
[0153] S3: Embedded predictive maintenance feedback, relying on the data accumulated in the digital twin in S1 to predict the device status, and feed the prediction results back to the dynamic optimization model in S2 to correct the candidate solution.
[0154] The S3 is embedded in the predictive maintenance feedback. The predictive maintenance module adopts a bidirectional long short-term memory network. The input data includes the equipment's historical operating parameters for the past 3 months and real-time status data for the past 72 hours. The output results include equipment fault type prediction, fault occurrence time interval, and remaining effective running time. The feedback method is to transmit the correction coefficient matrix to the dynamic optimization model.
[0155] The S3 embedded predictive maintenance feedback includes a correction process that recalculates the operating parameters of the equipment involved in the predicted failure in the candidate schemes, rearranges the timing of the coordinated actions of the associated equipment, and readjusts the material allocation scheme. After correction, the number of candidate schemes remains consistent with the initial number generated.
[0156] S4: Perform virtual-real interaction verification. Use the digital twin in S1 to perform virtual simulation of the candidate solutions modified in S3, select the optimal strategy and send it to the physical production line. At the same time, the physical production line data is continuously fed back to the digital twin in S1 to form a closed loop.
[0157] In the S4 virtual-real interaction verification, the virtual simulation verification includes single-factor variable testing, multi-condition combination testing, and extreme state simulation testing. The verification indicators include strategy execution time, total resource consumption, and number of products meeting quality standards. The selection rule is to determine the optimal strategy as the scheme with the highest comprehensive score.
[0158] In the S4 virtual-real interaction verification, the optimal strategy is distributed via an edge computing node relay, and the transmission protocol is MQTT. The physical production line feedback data includes the actual equipment parameters, actual production indicators, and actual quality test results after the strategy is executed. The closed-loop update cycle is once per hour.
[0159] S5: Construct a twin interaction platform. Based on the digital twin of S1, build a visual interactive interface to support manual intervention and adjustment of candidate solutions generated in S2, prediction results in S3, and verification process in S4. In the construction of the twin interaction platform in S5, the twin interaction platform adopts three-dimensional real-time rendering technology and supports gesture control and voice command input. The manual intervention includes adjusting the optimization target weight in S2, modifying the prediction threshold in S3, and pausing the strategy issuance process in S4.
[0160] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A dynamic optimization control method for industrial internet production lines based on digital twins, characterized in that, Includes the following steps; S1: Construct a synchronous mapping twin. Through real-time acquisition and fusion of multi-source heterogeneous data, generate a digital twin that maps to the physical production line and keep the two updated synchronously in real time. S2: Build a dynamic optimization model. Based on the digital twin constructed in S1, construct a control model that integrates deep reinforcement learning and adaptive fuzzy logic. Generate multiple sets of control strategy candidate schemes based on the preset optimization target. S3: Embedded predictive maintenance feedback, relying on the data accumulated in the digital twin in S1 to predict the device status, and feed the prediction results back to the dynamic optimization model in S2 to correct the candidate solution. S4: Perform virtual-real interaction verification. Use the digital twin in S1 to perform virtual simulation of the candidate solutions modified in S3, select the optimal strategy and send it to the physical production line. At the same time, the physical production line data is continuously fed back to the digital twin in S1 to form a closed loop. S5: Construct a twin interaction platform, build a visual interactive interface based on the digital twin of S1, and support manual intervention and adjustment of candidate solutions generated by S2, prediction results of S3 and verification process of S4.
2. The method for dynamic optimization control of industrial internet production lines based on digital twins according to claim 1, characterized in that, In the synchronous mapping twin constructed in S1, the fusion of multi-source heterogeneous data adopts an improved attention mechanism deep learning model. This model introduces the control strategy influence factor output by the dynamic optimization model in S2 as the attention weight adjustment parameter. By dynamically allocating the data fusion weight through the degree of influence of the control strategy on each link of the production line, the feature extraction accuracy of the data is improved.
3. The method for dynamic optimization control of industrial internet production lines based on digital twins according to claim 1, characterized in that, In the synchronous mapping twin constructed in S1, the multi-source heterogeneous data includes vibration values, temperature readings, and pressure parameters collected by equipment sensors; temperature and humidity indicators and dust concentration data obtained by environmental monitoring devices; material type, quantity, and conveying speed information recorded by the material management system; and appearance parameters, size data, and defect identifiers generated by product testing equipment.
4. The method for dynamic optimization control of industrial internet production lines based on digital twins according to claim 1, characterized in that, In the S2-based dynamic optimization model, deep reinforcement learning employs an improved DDPG algorithm that integrates the data fusion features from S1. The data features extracted from S1 are used as the input dimension of the state space, enabling the agent to prioritize data features that significantly affect the control effect during policy learning. At the same time, the real-time deviation of data fusion is included as a penalty in the reward function, enhancing the model's robustness to data synchronization delays.
5. The method for dynamic optimization control of industrial internet production lines based on digital twins according to claim 1, characterized in that, In the S2 dynamic optimization model, the adaptive fuzzy logic adapts to the time-varying characteristics of the production line by adjusting the membership function parameters online. The preset optimization targets include the production cycle time achievement rate, unit product energy consumption value, and product sampling pass rate. The working conditions simulated by the digital twin include material supply fluctuations, equipment load changes, and environmental parameter drift.
6. The method for dynamic optimization control of industrial internet production lines based on digital twins according to claim 1, characterized in that, The S3 is embedded in the predictive maintenance feedback. The predictive maintenance module adopts a bidirectional long short-term memory network. The input data includes the equipment's historical operating parameters for the past 3 months and real-time status data for the past 72 hours. The output results include equipment fault type prediction, fault occurrence time interval, and remaining effective running time. The feedback method is to transmit the correction coefficient matrix to the dynamic optimization model.
7. The method for dynamic optimization control of industrial internet production lines based on digital twins according to claim 1, characterized in that, The S3 embedded predictive maintenance feedback includes a correction process that recalculates the operating parameters of the equipment involved in the predicted failure in the candidate schemes, rearranges the timing of the coordinated actions of the associated equipment, and readjusts the material allocation scheme. After correction, the number of candidate schemes remains consistent with the initial number generated.
8. The method for dynamic optimization control of industrial internet production lines based on digital twins according to claim 1, characterized in that, In the S4 virtual-real interaction verification, the virtual simulation verification includes single-factor variable testing, multi-condition combination testing, and extreme state simulation testing. The verification indicators include strategy execution time, total resource consumption, and number of products meeting quality standards. The selection rule is to determine the optimal strategy based on the scheme with the highest comprehensive score.
9. The method for dynamic optimization control of industrial internet production lines based on digital twins according to claim 1, characterized in that, In the S4 execution of virtual-real interaction verification, the optimal strategy is distributed via an edge computing node, and the transmission protocol is MQTT. The physical production line feedback data includes the actual equipment parameters, actual production indicators, and actual quality test results after the strategy is executed. The closed-loop update cycle is once per hour.
10. The method for dynamic optimization control of industrial internet production lines based on digital twins according to claim 1, characterized in that, In the S5 twin interaction platform, the twin interaction platform adopts three-dimensional real-time rendering technology, supports gesture control and voice command input, and the manual intervention includes adjusting the optimization target weight in S2, modifying the prediction threshold in S3, and pausing the strategy issuance process in S4.