Steel pipe processing whole-process intelligent management and control method and system based on digital twinning

Abstract

The application discloses a steel pipe machining whole-process intelligent management and control method and system based on digital twinning, which comprises the following steps: collecting production process data, preprocessing to form a multi-source fusion data set; constructing a block twinning state vector, and establishing a twinning correlation graph; constructing an improved denoising diffusion implicit model to obtain a twinning state and state prediction; generating an abnormal evidence set, and constructing a restart vector; based on a restart random walk algorithm, obtaining a candidate ranking and path set; generating a cross-process linkage management and control strategy, and issuing and executing the strategy. Through the closed-loop cooperation of the improved denoising diffusion implicit model and the restart random walk algorithm, the application realizes the consistent state reconstruction, short-term prediction and linkage management and control of the whole process of steel pipe machining.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and in particular to a method and system for intelligent control of the entire steel pipe processing process based on digital twins. Background Technology

[0002] Steel pipe processing is a typical discrete-continuous hybrid manufacturing process, typically involving multiple sequential and parallel processing steps, collaborative operation of multiple equipment, and coupling of multiple inspection links. During production, multiple sources of data exist simultaneously, including process parameters, equipment operating status, quality inspection results, work order batch information, and logistics cycle time. Existing digital management and control solutions for steel pipe processing mostly adopt decentralized data acquisition and single-step control methods. Data sources exhibit inconsistencies in sampling frequency, time reference, identification systems, and data semantics. Furthermore, on-site data generally suffers from missing data, noise, drift, and detection lag, making it difficult to form a unified status representation for batches and processes. This results in a desynchronization between the digital twin model and the on-site status, insufficient reliability of the twin status, and difficulty in supporting real-time prediction, anomaly warning, and closed-loop management throughout the entire process.

[0003] Quality defects in steel pipe processing are often caused by the accumulation of multiple processes, fluctuations in equipment status, and the interaction of process parameters. The defect formation path spans multiple processes and equipment nodes, exhibiting strong coupling, strong temporal sequence, and strong correlation characteristics. Existing quality management and traceability solutions mostly rely on single-point inspection results or empirical rules for post-event judgment, lacking the ability to perform correlation modeling and cross-process reasoning for the entire process. This makes it difficult to achieve early warning of defects, root cause localization, and parameter linkage optimization. In traceability analysis, the common practice is to search and compare based on batch reports or single-process logs, lacking a unified expression of process sequence relationships, equipment affiliation relationships, parameter effect relationships, and inspection mapping relationships, making it difficult to reconstruct the cross-process impact chain. When quality fluctuations or concentrated defect outbreaks occur, existing solutions struggle to map anomalies from the phenomenological level to actionable targets in the parameter-equipment-process chain. Often, only stricter sampling inspections, line stoppages for investigation, or empirical parameter adjustments can be implemented, resulting in delayed handling and increased downtime losses. Furthermore, due to the lack of complete evidence chain records and explainable impact paths, it is difficult to form a closed loop for quality responsibility positioning and process improvement, affecting production stability, quality consistency, and delivery reliability.

[0004] Therefore, how to provide a method and system for intelligent control of the entire steel pipe processing process based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent control method and system for the entire steel pipe processing process based on digital twins. This invention comprehensively utilizes multi-source data fusion, twin association modeling, innovative denoising diffusion implicit model structure, and innovative restarted random walk structure methods to form a complete process control scheme encompassing production process data acquisition and preprocessing, block-based twin state construction and twin association graph establishment, twin state reconstruction and short-term prediction, anomaly evidence generation and transfer modeling, root cause tracing and impact path extraction, and cross-process linkage control strategy issuance, execution, and closed-loop updates. The improved denoising diffusion implicit model innovatively integrates a block coupling module, a graph prior injection module, and a feasible region projection module to achieve cross-block fusion denoising completion, relevant subspace focusing reconstruction, and process window and equipment capability boundary projection. The restarted random walk algorithm uses direct solution of steady-state equations and reverse elimination propagation to achieve core subgraph compression and back-substitution and root cause path output. Compared with existing technologies, this invention can improve the consistency between virtual and real synchronization and predictive availability, achieving early defect warning, cross-process root cause localization, and parameter linkage optimization. It possesses advantages such as strong interpretability, good real-time performance, and high engineering feasibility.

[0006] The intelligent control method for the entire steel pipe processing process based on digital twins according to embodiments of the present invention includes: Collect production process data from the entire steel pipe processing flow, preprocess the production process data, and form a multi-source fusion dataset; Based on the multi-source fusion dataset, construct block-based twin state vectors according to multi-dimensional business domains, and establish a twin association graph; An improved denoising diffusion implicit model is constructed. Based on the block coupling module, cross-block fusion denoising and missing completion are performed on the block twin state vector. The graph prior injection module is introduced to perform relevant subspace focusing reconstruction on the twin association graph. The feasible region projection module is used to project the process window and equipment capacity boundary to obtain the twin state reconstruction results and short-term state prediction results. Based on the twin state reconstruction results and short-term state prediction results, an abnormal evidence set is generated, a restart vector is constructed, and the twin association graph is converted into a transition matrix; Based on the restart random walk algorithm, the steady-state equations of the transition matrix and restart vector are directly solved. Core subgraph compression and back-substitution calculation are performed through reverse elimination propagation to obtain the root cause candidate ranking and the set of influencing paths. By ranking root cause candidates and generating a set of impact paths, a cross-process linkage control strategy is generated, issued and executed, and the twin status, twin relationship diagram and abnormal evidence set are updated.

[0007] Optionally, the production process data includes process parameter data, equipment operation data, quality inspection data, work order batch data, and logistics cycle time data.

[0008] Optionally, forming the multi-source fusion dataset includes: Production process data is collected from each step of the entire steel pipe processing process, and batch identifier, process identifier, equipment identifier and timestamp are generated for each data record; The production process data is uniformly encoded and time-aligned. Data with different sampling frequencies are mapped to the same time axis according to the timestamp. Data with multiple records corresponding to the same time axis position are merged, and data with no corresponding record at the time axis position are marked as missing. The time-aligned production process data is preprocessed to form a multi-source fusion dataset. The preprocessing includes unifying units and dimensions, removing outliers and recording removal marks, marking missing fields and recording missing position indexes, and aggregating the data by batch identifier and process identifier to generate the multi-source fusion dataset.

[0009] Optionally, the construction of the block-based twin state vector and the establishment of the twin association graph include: Based on the multi-source fusion dataset, a block-based twin state vector is generated according to the batch identifier and process identifier. The block-based twin state vector includes process block, equipment block, quality block, work order batch block and logistics cycle block. Each block is formed by concatenating the corresponding fields in dimensional order. Missing fields in each block are represented by the position code corresponding to the missing marker. The process node set is determined based on the process identifier, the equipment node set is determined based on the equipment identifier, the parameter node set is determined based on the process parameter field, the quality node set is determined based on the quality inspection field, the defect node set is determined based on the defect category field, and the batch node set is determined based on the batch identifier. The node sets are then merged to form the node set of the twin association diagram. Based on the multi-source fusion dataset and the block twin state vector, an edge set of twin association graph is generated. The edge set includes process sequence edges, equipment membership edges, parameter effect edges, detection mapping edges, and historical co-occurrence edges.

[0010] Optionally, obtaining the twin state reconstruction results and short-term state prediction results includes: An improved denoising diffusion implicit model is constructed, which includes a diffusion process module, a denoising module, a block coupling module, a graph prior injection module, and a feasible region projection module. The diffusion process module generates a forward noise-adding sequence and a reverse noise-de-noising sequence. The forward noise-adding sequence covers time step zero to the end of time step, and generates a noisy state vector for each time step. The noisy state vector is obtained by linear combination of the original state vector and the standard normal noise vector. The coefficient of the original state vector is the square root of the cumulative hold coefficient, and the coefficient of the standard normal noise vector is the square root minus the cumulative hold coefficient. The cumulative hold coefficient is determined by the noise schedule. The reverse noise-de-noising sequence covers time step end to time step zero, and updates the noisy state vector step by step. The block coupling module performs cross-block fusion denoising and missing completion processing on the block twin state vectors. It encodes the process block, equipment block, quality block, work order batch block and logistics cycle block to obtain block representation vectors. Based on the inter-block interaction mapping, the block representation vectors are synthesized into a fused representation vector. The graph prior injection module generates graph prior vectors based on the Siamese association graph, and performs dimension selection operations on the fused representation vectors to obtain focused representation vectors; The denoising module takes the noisy state vector, time step and focused representation vector as input, and outputs the noise estimation vector. Based on the inverse solution of the noise estimation vector, the reconstruction estimation vector of the original state vector is obtained, forming the twin state reconstruction result and the short-term state prediction result. The feasible region projection module performs process window and equipment capacity boundary projection processing on the twin state reconstruction results and short-term state prediction results.

[0011] Optionally, the construction of the restart vector, which converts the twin association graph into a transition matrix, includes: An abnormal evidence set is generated based on the twin state reconstruction results and short-term state prediction results. The abnormal evidence set is obtained by comparing process parameter values ​​with the upper and lower bounds of the process window, comparing equipment state values ​​with equipment capacity boundaries, comparing quality index values ​​with quality threshold ranges, matching defect category fields, and comparing logistics cycle time values ​​with target cycle time ranges. A restart vector is constructed based on the set of abnormal evidence. The dimension of the restart vector is consistent with the number of nodes in the Siamese graph. The node position corresponding to the abnormal evidence is assigned a value of one, and the remaining node positions are assigned a value of zero. The normalized restart vector is obtained by dividing each dimension value by the sum of the dimension values. The twin graph is converted into a transition matrix. The dimension of the transition matrix is ​​the same as the number of nodes in the twin graph. When there is an edge, the element takes the edge weight; when there is no edge, the element takes zero. The row-normalized transition matrix is ​​obtained by dividing each row element by the sum of the elements in the current row.

[0012] Optionally, obtaining the root cause candidate ranking and the set of influencing paths includes: Input the transition matrix, restart vector, and restart probability to construct the full node correlation distribution vector and determine the identity matrix with the same dimension as the transition matrix; The steady-state equation is solved directly. The steady-state equation satisfies the condition that the product of the identity matrix minus the product of the restart probability and the transpose of the transition matrix, multiplied by the correlation distribution vector, equals the product of the restart probability and the restart vector. The correlation distribution vector of all nodes is obtained by solving the sparse linear equation system. Core subgraph compression and back-substitution calculation are performed through reverse elimination propagation. Core subgraph compression includes determining the set of eliminateable nodes and the set of retained nodes based on the sequence of processes and the affiliation of equipment. Node elimination is performed on the set of eliminateable nodes to obtain the core subgraph transition matrix and generate the back-substitution mapping relationship. Back-substitution calculation includes calculating the correlation distribution of retained nodes on the core subgraph and restoring the correlation distribution of eliminateable nodes based on the back-substitution mapping relationship. The root cause candidate ranking and the set of influencing paths are generated based on the correlation distribution vector of all nodes. The root cause candidate ranking is obtained by sorting the correlation distribution vectors from largest to smallest value. The set of influencing paths is determined by the set of reachable paths from the node corresponding to the restart vector to the root cause candidate node in the core subgraph transition matrix, and is obtained by sorting the cumulative values ​​of the correlation distribution vectors on the path from largest to smallest value.

[0013] Optionally, the generation and execution of the cross-process linkage control strategy includes: The root cause candidate node set is determined based on the root cause candidate ranking, and the process node, equipment node, parameter node and quality node set corresponding to the root cause candidate node set is determined based on the influence path set. Cross-process linkage control strategies are generated by combining process nodes, equipment nodes, parameter nodes, and quality nodes. These strategies include process parameter adjustment instructions, equipment operation adjustment instructions, stricter inspection instructions, and process handling instructions. Issue cross-process linkage control strategies and collect execution results. The execution results include execution process identifiers, execution equipment identifiers, execution parameter settings, and execution detection results. Write the execution results into the twin status and update the node status, edge set, and abnormal evidence set of the twin association graph based on the execution results.

[0014] The intelligent control system for the entire steel pipe processing process based on digital twins according to embodiments of the present invention includes the following modules: The data acquisition and preprocessing module is used to collect data from the entire steel pipe processing production process and complete preprocessing to form a multi-source fusion dataset. The twin state construction module is used to construct block twin state vectors based on multi-source fusion datasets and establish twin association graphs. The twin reconstruction prediction module is used to construct an improved denoised diffusion implicit model, and generates twin state reconstruction results and prediction results through block twin state vectors and twin association graphs; The Anomaly Evidence and Transition Modeling Module is used to generate anomaly evidence sets and construct restart vectors based on reconstruction and prediction results, and to convert twin correlation graphs into transition matrices. The root cause tracing calculation module is used to complete the direct solution and reverse elimination propagation of the steady-state equation based on the restart random walk algorithm, according to the transition matrix and restart vector, and output the root cause candidate ranking and the set of influence paths; The linkage control execution module is used to generate cross-process linkage control strategies based on the root cause candidate ranking and the set of impact paths, and to issue and execute them, and update the twin status, twin relationship diagram and abnormal evidence set.

[0015] The beneficial effects of this invention are: This invention preprocesses and fuses data from the entire steel pipe processing production process to construct segmented twin state vectors and twin correlation graphs. It employs an improved denoising diffusion implicit model to achieve cross-block fusion denoising and missing data completion, focused reconstruction of related subspaces, and projection of process windows and equipment capability boundaries. This ensures that the digital twin state remains consistent with the actual on-site state and possesses short-term predictive capabilities. Compared to control methods based on single-process modeling or rule thresholds, this invention improves the stability and usability of twin state reconstruction, enhances the timeliness of anomaly identification and risk warning, and outputs states and suggestions that meet process executable boundaries, reducing the risk of misjudgment and miscontrol, even under conditions of multi-source heterogeneity, asynchronous sampling, missing data, and noise interference.

[0016] This invention constructs a restart vector and generates a transition matrix based on a set of abnormal evidence. It employs a restart random walk algorithm to obtain the correlation distribution of all nodes through direct solution of the steady-state equation, and uses reverse elimination propagation to achieve core subgraph compression and back-substitution calculation, thereby outputting a root cause candidate ranking and a set of influencing paths. This generates a cross-process linkage control strategy for deployment and closed-loop updates. Compared to methods relying on post-event detection or manual experience-based traceability, this invention enables interpretable traceability and responsibility location across the process-quality coupling link, supports collaborative decision-making for parameter linkage optimization, stricter detection, and process handling, reduces downtime investigation and rework costs, and improves quality consistency, production stability, and delivery reliability. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the intelligent control method for the entire steel pipe processing process based on digital twins proposed in this invention; Figure 2 This is a structural block diagram of the improved noise reduction and diffusion implicit model of the intelligent control method for the entire steel pipe processing process based on digital twins proposed in this invention. Figure 3 This is a functional diagram of the intelligent control system for the entire steel pipe processing process based on digital twins proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 and Figure 2 A digital twin-based intelligent control method for the entire steel pipe processing process includes: Collect production process data from the entire steel pipe processing flow, preprocess the production process data, and form a multi-source fusion dataset; Based on the multi-source fusion dataset, construct block-based twin state vectors according to multi-dimensional business domains, and establish a twin association graph; An improved denoising diffusion implicit model is constructed. Based on the block coupling module, cross-block fusion denoising and missing completion are performed on the block twin state vector. The graph prior injection module is introduced to perform relevant subspace focusing reconstruction on the twin association graph. The feasible region projection module is used to project the process window and equipment capacity boundary to obtain the twin state reconstruction results and short-term state prediction results. Based on the twin state reconstruction results and short-term state prediction results, an abnormal evidence set is generated, a restart vector is constructed, and the twin association graph is converted into a transition matrix; Based on the restart random walk algorithm, the steady-state equations of the transition matrix and restart vector are directly solved. Core subgraph compression and back-substitution calculation are performed through reverse elimination propagation to obtain the root cause candidate ranking and the set of influencing paths. By ranking root cause candidates and generating a set of impact paths, a cross-process linkage control strategy is generated, issued and executed, and the twin status, twin relationship diagram and abnormal evidence set are updated.

[0020] In this embodiment, the production process data includes process parameter data, equipment operation data, quality inspection data, work order batch data, and logistics cycle time data.

[0021] In this embodiment, forming a multi-source fusion dataset includes: Production process data is collected from each step of the entire steel pipe processing process, and batch identifier, process identifier, equipment identifier and timestamp are generated for each data record; The production process data is uniformly encoded and time-aligned. Data with different sampling frequencies are mapped to the same time axis according to the timestamp. Data with multiple records corresponding to the same time axis position are merged, and data with no corresponding record at the time axis position are marked as missing. The time-aligned production process data is preprocessed to form a multi-source fusion dataset. The preprocessing includes unifying units and dimensions, removing outliers and recording removal marks, marking missing fields and recording missing position indexes, and aggregating the data by batch identifier and process identifier to generate the multi-source fusion dataset.

[0022] In this embodiment, the construction of block-based twin state vectors and the establishment of twin association graphs include: Based on the multi-source fusion dataset, a block-based twin state vector is generated according to the batch identifier and process identifier. The block-based twin state vector includes process block, equipment block, quality block, work order batch block and logistics cycle block. Each block is formed by concatenating the corresponding fields in dimensional order. Missing fields in each block are represented by the position code corresponding to the missing marker. The process node set is determined based on the process identifier, the equipment node set is determined based on the equipment identifier, the parameter node set is determined based on the process parameter field, the quality node set is determined based on the quality inspection field, the defect node set is determined based on the defect category field, and the batch node set is determined based on the batch identifier. The node sets are then merged to form the node set of the twin association diagram. An edge set for the twin association graph is generated based on the multi-source fusion dataset and the segmented twin state vectors. This edge set includes process sequence edges, equipment affiliation edges, parameter effect edges, detection mapping edges, and historical co-occurrence edges. Specifically, the edge set for generating the twin association graph based on the multi-source fusion dataset and the segmented twin state vectors is as follows: The multi-source fusion dataset is aggregated by batch identifier and the process records are sorted by timestamp. Process sequence edges are established between adjacent process nodes. Equipment membership edges are established between process nodes and equipment nodes according to the correspondence between process identifier and equipment identifier. The edge weight is the proportion of the occurrence of the correspondence to the total number of process records. Within the same batch and the same process, process parameter fields and quality inspection fields are extracted based on block twin state vectors. Pearson correlation coefficient is calculated for each pair of parameter nodes and quality nodes. When the absolute value of the Pearson correlation coefficient is not less than 0.6, parameter action edges are established and the edge weight is the absolute value of the correlation coefficient. Calculate the co-occurrence conditional probability of the quality inspection field and the defect category field. When the co-occurrence conditional probability is not less than 0.3, establish a detection mapping edge between the quality node and the defect node, and take the co-occurrence conditional probability as the edge weight. Calculate the co-occurrence ratio of cross batches under the same equipment identifier or the same defect category. When the co-occurrence ratio is not less than 0.2, establish a historical co-occurrence edge between batch nodes, and take the co-occurrence ratio as the edge weight.

[0023] In this embodiment, obtaining the twin state reconstruction result and the short-term state prediction result includes: An improved denoising diffusion implicit model is constructed, which includes a diffusion process module, a denoising module, a block coupling module, a graph prior injection module, and a feasible region projection module. Specifically, the construction of the improved denoising diffusion implicit model is as follows: The block coupling module is embedded into the input encoding position of the original denoising module. The block encoding and inter-block interaction mapping of process block, equipment block, quality block, work order batch block and logistics cycle block are used to replace the single vector encoding. The graph prior injection module is connected between the block coupling output and the input of the denoising module. The graph prior vector generated by the twin association graph is used to perform dimension selection operation to form a focused representation vector and participate in the noise estimation calculation. After the denoising module outputs the reconstructed estimation vector, the feasible region projection module is connected in series. The upper and lower bounds of the process window and the equipment capacity boundary are used to perform interval projection on the reconstructed estimation vector to obtain the executable twin state, thus forming an improved denoising diffusion implicit model. The diffusion process module generates a forward denoising sequence and a reverse denoising sequence. The forward denoising sequence covers time step zero to the end of time step, generating a noisy state vector for each time step. The noisy state vector is obtained by a linear combination of the original state vector and the standard normal noise vector. The coefficients of the original state vector are the square root of the cumulative hold coefficient, and the coefficients of the standard normal noise vector are the square root minus the cumulative hold coefficient, which is determined by the noise schedule. The reverse denoising sequence covers time step end to time step zero and updates the noisy state vector step by step. Specifically, the diffusion process module generates the forward denoising sequence and the reverse denoising sequence as follows: The diffusion process module first sets the end time step to 1000 and sets the noise level to linearly increasing. The noise intensity of the first step is 0.001 and the noise intensity of the last step is 0.02. For each time step, the retention coefficient is calculated by subtracting the noise intensity of the step from one. The retention coefficients from the first step to the current step are multiplied sequentially to obtain the cumulative retention coefficient. When generating the forward-noise sequence, at each time step, a noise vector with the same dimension as the original state vector is sampled from the standard normal distribution. The original state vector and the noise vector are linearly combined to obtain the noisy state vector. The coefficient of the original state vector is the square root of the cumulative retention coefficient, and the coefficient of the noise vector is the square root of the cumulative retention coefficient minus one, resulting in a noisy state vector sequence from time step zero to the end of the time step. When generating the reverse denoising sequence, starting from the noisy state vector at the end of the time step, the sampling step number is set to 20. The reverse time step sequence is selected at equal intervals. The current noisy state vector is input to the denoising module at the corresponding time step to obtain the noise estimation vector. The reconstructed estimation vector is obtained by inverse solving the noise estimation vector. The noisy state vector of the next time step is updated until the output of time step zero is obtained, forming the reverse denoising sequence from the end of the time step to time step zero. The block coupling module performs cross-block fusion denoising and missing word completion processing on the block twin state vectors. It encodes the process block, equipment block, quality block, work order batch block, and logistics cycle time block to obtain block representation vectors. Based on inter-block interaction mapping, the block representation vectors are synthesized into a fused representation vector. Specifically, the block coupling module performs cross-block fusion denoising and missing word completion processing on the block twin state vectors as follows: The block coupling module splits the block twin state vector into process block, equipment block, quality block, work order batch block, and logistics cycle block, and inputs them into the corresponding encoders. For each block, the current block feature vector is first constructed according to the field dimension order, and the missing positions are set to 0 while generating a missing mask vector. Each encoder consists of two fully connected mapping layers. The first layer maps the current block feature vector to a 128-dimensional latent vector and uses non-linear activation. The second layer maps the 128-dimensional latent vector to a 64-dimensional block representation vector. The inter-block interaction mapping concatenates the five block representation vectors in sequence into a 320-dimensional vector. A 256-dimensional fusion representation vector is obtained through one fully connected mapping layer. At the same time, based on the missing mask vector, the components of the corresponding missing fields in the fusion representation vector are marked as to be filled. The reconstructed estimation vector output by the denoising module is backfilled according to the original field position. The position with a missing mask of 1 is replaced with the reconstructed estimation value, and the position with a missing mask of 0 retains the original observation value, thus completing cross-block fusion denoising and missing completion. The graph prior injection module generates graph prior vectors based on the Siamese relational graph, and performs dimension selection operations on the fused representation vectors to obtain focused representation vectors, where: Generate graph prior vectors, specifically: The twin graph is converted into a row-normalized transition matrix. The restart vector and restart probability 0.15 are input, and the steady-state equation is solved to obtain the correlation distribution of all nodes. The correlation distribution of all nodes is sorted from largest to smallest, and the top 5% of nodes are selected as the set of highly correlated nodes. A mapping table from graph nodes to the dimensions of the block twin state vector is established. The state dimension positions corresponding to the set of highly correlated nodes are set to 1, and the remaining positions are set to 0, to obtain the binary graph prior vector with the same dimension as the fused representation vector. The sum of the binary graph prior vectors after summing by length is normalized so that the sum of the values ​​of each dimension is 1. Solving the steady-state equation means that, under the condition that the unknown quantity is the correlation distribution vector of all nodes, a linear system of equations is established. The identity matrix minus 1 minus the product of the restart probability and the transpose of the transition matrix, and then multiplied by the correlation distribution vector, is equal to the product of the restart probability and the restart vector. The fusion representation vector is obtained by performing a dimension selection operation on the fusion representation vector. Specifically, the fusion representation vector is multiplied dimension by dimension of the binary map prior vector to obtain the focused representation vector. The dimension of the binary map prior vector that is 0 is set to 0, and the dimension of the fusion representation vector that is 1 retains the value of the dimension of the fusion representation vector. The denoising module takes the noisy state vector, time step, and focused representation vector as input, and outputs a noise estimation vector. Based on the inverse solution of the noise estimation vector, a reconstructed estimation vector of the original state vector is obtained, forming a twin state reconstruction result and a short-term state prediction result, wherein: The output noise estimation vector is obtained as follows: the denoising module concatenates the noisy state vector and the focused representation vector according to their dimensions to form the input vector. The time step is encoded into a time step embedding vector of the same length as the input vector and then added to the input vector. The noise estimation vector is obtained by passing through three fully connected layers in sequence. The first layer maps the input vector to 512 dimensions and applies nonlinear activation. The second layer maps the 512 dimensions to 512 dimensions and applies nonlinear activation. The third layer maps the 512 dimensions to the same dimension as the noisy state vector and outputs the noise estimation vector. The noise estimation vector is truncated element by element to the interval [-3,3] to suppress abnormal amplification. Based on the inverse solution of the noise estimation vector, the reconstructed estimation vector of the original state vector is obtained, forming the twin state reconstruction result and the short-term state prediction result, specifically: Given the noisy state vector, cumulative preservation coefficient, and noise estimation vector at the time step, the reconstruction estimate based on the original state vector is equal to the noisy state vector minus the square root of the cumulative preservation coefficient multiplied by the noise estimation vector, and then divided by the square root of the cumulative preservation coefficient to obtain the reconstruction estimate vector. The reconstruction estimate vector is then projected through the feasible region to obtain the twin state reconstruction result. Using the twin state reconstruction result as the starting state, a noisy state vector is generated by forward noise addition, and then reverse noise reduction and update are performed. This process is repeated 5 times to obtain 5 steps of short-term state prediction results. The short-term state prediction results are arranged in time order to form a short-term state prediction sequence. The feasible region projection determines the upper and lower bounds of the process parameters and equipment capabilities for each dimension based on the batch identifier and process identifier. The value of each dimension is truncated. If the current dimension value is less than the lower bound, the current dimension value is replaced with the lower bound. If the current dimension value is greater than the upper bound, the current dimension value is replaced with the upper bound. If the current dimension value is between the lower and upper bounds, it remains unchanged. Dimensions belonging to logical or enumerated fields are mapped according to the nearest valid value. The value set is determined by the encoding table of the work order batch field and the process disposal field. The vector after dimension-by-dimensional processing is used as the feasible region projection output and as the input for the twin state reconstruction result and short-term state prediction. The feasible region projection module performs process window and equipment capacity boundary projection processing on the twin state reconstruction results and short-term state prediction results.

[0024] In this embodiment, constructing the restart vector to convert the twin association graph into a transition matrix includes: An anomaly evidence set is generated based on the twin state reconstruction results and short-term state prediction results. This set is obtained by comparing process parameter values ​​with the upper and lower bounds of the process window, comparing equipment state values ​​with equipment capacity boundaries, comparing quality index values ​​with quality threshold ranges, matching defect category fields, and comparing logistics cycle time values ​​with target cycle time ranges. Specifically, the generation of the anomaly evidence set includes: Based on the twin state reconstruction results and short-term state prediction results, process parameters, equipment status, quality indicators, defect categories and logistics cycle time fields are extracted one by one according to batch identifier and process identifier. The lower and upper boundaries of the process window, the lower and upper boundaries of the capacity boundary of the corresponding equipment, the lower and upper boundaries of the threshold of the corresponding quality indicator and the lower and upper boundaries of the target cycle time are read from the pre-stored table. When the process parameter value is less than the lower boundary of the process window or greater than the upper boundary of the process window, process over-limit evidence is generated and the over-limit range is recorded as the difference between the current value and the nearest boundary value. When the equipment status value is less than the lower boundary of the capacity boundary or greater than the upper boundary of the capacity boundary, equipment anomaly evidence is generated and the deviation range is recorded. When the quality index value is less than the lower threshold of 0.95 or greater than the upper threshold of 1.05, quality exceedance evidence is generated and the deviation range is recorded. When the defect category field matches the preset defect category set, defect hit evidence is generated. The preset defect category set is configured by the system and includes the defect category codes that need to be warned. When the logistics cycle time value is less than the lower limit of the target cycle time or greater than the upper limit of the target cycle time, cycle time deviation evidence is generated and the deviation range is recorded. All kinds of evidence generated in the same batch and the same process are summarized into an abnormal evidence set according to evidence type, corresponding node identifier and deviation range. Evidence with an absolute value of deviation range greater than or equal to 0.05 is marked as valid evidence. A restart vector is constructed based on the set of abnormal evidence. The dimension of the restart vector is consistent with the number of nodes in the Siamese graph. The node position corresponding to the abnormal evidence is assigned a value of one, and the remaining node positions are assigned a value of zero. The normalized restart vector is obtained by dividing each dimension value by the sum of the dimension values. The twin graph is converted into a transition matrix. The dimension of the transition matrix is ​​the same as the number of nodes in the twin graph. When there is an edge, the element takes the edge weight; when there is no edge, the element takes zero. The row-normalized transition matrix is ​​obtained by dividing each row element by the sum of the elements in the current row.

[0025] In this embodiment, obtaining the root cause candidate ranking and the set of influencing paths includes: Given the transition matrix, restart vector, and restart probability, construct the full node correlation distribution vector and determine the identity matrix with the same dimension as the transition matrix. Specifically, the construction of the full node correlation distribution vector is as follows: The number of nodes in the transition matrix is ​​determined as the dimension length of the correlation distribution vector. The normalized restart vector is used as the initial value of the correlation distribution vector, and the restart probability is set to 0.15. In each solution iteration, the new correlation distribution vector is updated by subtracting the product of the restart probability and the transpose of the transition matrix multiplied by the previous correlation distribution vector, and then adding the product of the restart probability and the restart vector. This process continues until the sum of the absolute values ​​of the differences in each dimension of the correlation distribution vectors of two adjacent iterations is less than or equal to 1 × 10 to the power of -6. Finally, the correlation distribution vector of all nodes is obtained. The identity matrix of the same dimension is defined as a square matrix with diagonal elements of 1 and off-diagonal elements of 0. The steady-state equation is solved directly. The steady-state equation satisfies the condition that the product of the identity matrix minus the product of the restart probability and the transpose of the transition matrix, multiplied by the correlation distribution vector, equals the product of the restart probability and the restart vector. The correlation distribution vector of all nodes is obtained by solving the sparse linear equation system. Core subgraph compression and back-substitution calculation are performed through reverse elimination propagation. Core subgraph compression includes determining the set of eliminateable nodes and the set of retained nodes based on the sequence of processes and equipment affiliation; performing node elimination on the set of eliminateable nodes to obtain the core subgraph transition matrix; and generating a back-substitution mapping relationship. Back-substitution calculation includes calculating the correlation distribution of retained nodes on the core subgraph and restoring the correlation distribution of eliminateable nodes based on the back-substitution mapping relationship. To obtain the core subgraph transition matrix, node elimination is performed on the set of eliminateable nodes. Specifically, in the twin association graph, the set of retained nodes is rearranged to the front by index, and the set of eliminateable nodes is rearranged to the back. The original transition matrix is ​​divided into 4 sub-blocks in sequence. When performing node elimination on the set of eliminateable nodes, the product of the identity matrix minus 1 minus the restart probability 0.15 and the transpose of the eliminateable sub-block is first constructed to obtain the eliminateable coefficient sub-block. The product of the retained to eliminateable coupling sub-block and the inverse of the eliminateable coefficient sub-block is calculated. Finally, the product is used to eliminate and update the eliminateable to retained coupling sub-block to obtain the equivalent core subgraph coefficient matrix containing only retained nodes. The core subgraph transition matrix is ​​obtained by subtracting the equivalent core subgraph coefficient matrix from the identity matrix. The core subgraph transition matrix is ​​kept sparsely stored, and row normalization is performed on each row so that the sum of the row elements is 1. Generate back-substitution mapping relationships. Back-substitution calculation includes calculating the correlation distribution of retained nodes on the core subgraph, and recovering the correlation distribution of reducible nodes based on the back-substitution mapping relationships. Specifically: The back-substitution mapping relationship is composed of the inverse of the eliminateable coefficient sub-block obtained in the elimination process multiplied by the eliminateable constant sub-vector, and the inverse of the eliminateable coefficient sub-block multiplied by the eliminateable to retain coupling sub-block. The former records the baseline contribution of the eliminateable node under the only drive of the restart vector, and the latter records the linear response coefficient of the eliminateable node to the correlation distribution of the retainable node. When calculating the correlation distribution of retained nodes on the core subgraph, the correlation distribution vector of retained nodes is obtained by solving the core subgraph transition matrix and core subgraph restart vector using the direct solution of the steady-state equation. When back-substituting to recover the correlation distribution of reducible nodes, the correlation distribution vector of retained nodes is substituted into the back-substituting mapping relationship. The correlation distribution vector of reducible nodes is calculated as the baseline contribution vector plus the linear response coefficient matrix multiplied by the correlation distribution vector of retained nodes. The correlation distribution of reducible nodes is then merged according to the index position before rearrangement and normalized by sum so that the sum of the values ​​of each dimension is 1. The root cause candidate ranking and the set of influencing paths are generated based on the correlation distribution vector of all nodes. The root cause candidate ranking is obtained by sorting the correlation distribution vectors from largest to smallest value. The set of influencing paths is determined by the set of reachable paths from the node corresponding to the restart vector to the root cause candidate node in the core subgraph transition matrix, and is obtained by sorting the cumulative values ​​of the correlation distribution vectors on the path from largest to smallest value.

[0026] In this embodiment, the generation and execution of the cross-process linkage control strategy includes: The root cause candidate node set is determined based on the root cause candidate ranking, and the process node, equipment node, parameter node and quality node set corresponding to the root cause candidate node set is determined based on the influence path set. Cross-process linkage control strategies are generated by combining process nodes, equipment nodes, parameter nodes, and quality nodes. These strategies include process parameter adjustment instructions, equipment operation adjustment instructions, stricter inspection instructions, and process handling instructions. Issue cross-process linkage control strategies and collect execution results. The execution results include execution process identifiers, execution equipment identifiers, execution parameter settings, and execution detection results. Write the execution results into the twin status and update the node status, edge set, and abnormal evidence set of the twin association graph based on the execution results.

[0027] refer to Figure 3 The intelligent control system for the entire steel pipe processing process based on digital twins includes the following modules: The data acquisition and preprocessing module is used to collect data from the entire steel pipe processing production process and complete preprocessing to form a multi-source fusion dataset. The twin state construction module is used to construct block twin state vectors based on multi-source fusion datasets and establish twin association graphs. The twin reconstruction prediction module is used to construct an improved denoised diffusion implicit model, and generates twin state reconstruction results and prediction results through block twin state vectors and twin association graphs; The Anomaly Evidence and Transition Modeling Module is used to generate anomaly evidence sets and construct restart vectors based on reconstruction and prediction results, and to convert twin correlation graphs into transition matrices. The root cause tracing calculation module is used to complete the direct solution and reverse elimination propagation of the steady-state equation based on the restart random walk algorithm, according to the transition matrix and restart vector, and output the root cause candidate ranking and the set of influence paths; The linkage control execution module is used to generate cross-process linkage control strategies based on the root cause candidate ranking and the set of impact paths, and to issue and execute them, and update the twin status, twin relationship diagram and abnormal evidence set.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous production line of a steel pipe processing company. The process chain covers multiple steps including forming, welding, heat treatment, straightening, cutting, and non-destructive testing. On-site issues included discrepancies between process settings and actual measurements, equipment status fluctuations, delayed test results, and data asynchrony caused by different sampling frequencies and clock drift. Furthermore, some sensor points experienced intermittent loss after replacement and maintenance. In the current scenario, although a digital twin system has been deployed, the twin's state frequently deviates from the actual on-site state. Quality defects are often only exposed in subsequent testing stages, making it difficult to trace back to key parameters across processes and equipment root causes. This results in anomaly handling relying primarily on experience-based troubleshooting, leading to high downtime and re-inspection costs and an incomplete quality responsibility chain.

[0029] After integrating the method of this invention into the production line data bus, the production process data first undergoes unified encoding, time alignment, missing value marking, and outlier removal to form a multi-source fusion dataset oriented towards batches and processes. Subsequently, segmented twin state vectors are constructed according to business domains such as process, equipment, quality, batch, and cycle time, and twin association graphs are constructed based on process sequence, equipment affiliation, parameter effects, detection mapping, and historical co-occurrence relationships. The improved denoising diffusion implicit model performs cross-block fusion denoising and missing value completion on the segmented states in the segmented coupling module, focuses and reconstructs the highly correlated subspace given by the association graph in the graph prior injection module, and then projects the reconstruction results onto the process window and equipment capability boundary by the feasible region projection module to obtain executable twin state reconstruction results and short-term prediction results. An abnormal evidence set is generated from the prediction and reconstruction results, and a restart vector is constructed. The correlation graph is converted into a transition matrix. The random walk algorithm is restarted to obtain the correlation distribution of all nodes through the direct solution of the steady-state equation. Then, the core subgraph is compressed and back-substituted through reverse elimination propagation. The root cause candidate ranking and the set of influence paths are output. Finally, a cross-process linkage control strategy is generated and issued for execution. The execution results are written back to update the twin state, correlation graph and abnormal evidence set.

[0030] During continuous operation and evaluation cycles, the system maintains stable updates to the twin state under complex conditions of multi-specification switching and equipment maintenance. This significantly improves the issues of data gaps, noise, and asynchrony-induced discrepancies between virtual and real data, ensuring the twin results have the credibility and executability for real-time control. For typical quality issues such as weld defects and dimensional deviations, the system can provide risk warnings before defects become explicit and generate combined strategies for parameter adjustment, stricter inspection, and process handling. After on-site implementation, quality fluctuations are suppressed, and re-inspection pass rates are more stable. In terms of root cause tracing, the output root cause candidates and impact paths narrow the investigation scope to combinations of key equipment and key parameters, improving the efficiency of anomaly location and handling. At the same time, it forms a complete evidence chain of batch—process—equipment—parameter—inspection—handling, facilitating quality responsibility identification and continuous process improvement, reducing the risk of rework and scrap, and improving production stability and delivery consistency.

[0031] Table 1. Comprehensive Comparison of Intelligent Management and Control Solutions for the Entire Steel Pipe Processing Process

[0032] As shown in Table 1, in terms of data synchronization and data quality, the state synchronization error of the present invention is 4.8 seconds, which is significantly lower than the manufacturing execution threshold of 95.0 seconds, single-process statistical control of 72.0 seconds, cyclic prediction of 40.0 seconds, auto-encoding anomaly of 55.0 seconds, and pure graph tracing of 68.0 seconds. In terms of completion mean square error, the present invention has an error of 0.012, which is also lower than the cyclic prediction of 0.034, auto-encoding anomaly of 0.028, single-process statistical control of 0.061, pure graph tracing of 0.079, and manufacturing execution threshold of 0.085, indicating that the overall error is the smallest in multi-source fusion and missing completion.

[0033] In terms of early warning and identification effectiveness, the early warning lead time of this invention is 32 minutes, which is higher than the 18 minutes of cyclic prediction, the 12 minutes of self-encoded anomaly, the 8 minutes of single-process statistical control, and the 5 minutes of pure drawing traceability, while the manufacturing execution threshold is 0 minutes. Regarding the defect detection rate, this invention achieves 96.2%, which is higher than the 90.3% of self-encoded anomaly, the 88.1% of cyclic prediction, the 86.0% of pure drawing traceability, the 83.4% of single-process statistical control, and the 78.5% of the manufacturing execution threshold. Meanwhile, the false alarm rate is 4.1%, lower than the 7.5% of cyclic prediction, the 6.9% of self-encoded anomaly, the 9.8% of single-process statistical control, the 10.7% of pure drawing traceability, and the 12.6% of the manufacturing execution threshold.

[0034] In terms of traceability, location, and handling efficiency, the root cause hit rate of the present invention is 92.4%, which is higher than that of pure map traceability (80.3%), cyclic prediction (60.1%), self-encoded anomaly (58.7%), single-process statistical control (52.8%), and manufacturing execution threshold (45.2%). The strategy execution success rate is 98.1%, which is higher than that of cyclic prediction (93.2%), self-encoded anomaly (92.5%), single-process statistical control (91.0%), pure map traceability (90.6%), and manufacturing execution threshold (89.4%). The average handling time is 16 minutes, which is significantly shorter than that of pure map traceability (35 minutes), cyclic prediction (44 minutes), self-encoded anomaly (48 minutes), single-process statistical control (62 minutes), and manufacturing execution threshold (85 minutes).

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent control of the entire steel pipe processing process based on digital twins, characterized in that, include: Collect production process data from the entire steel pipe processing flow, preprocess the production process data, and form a multi-source fusion dataset; Based on the multi-source fusion dataset, construct block-based twin state vectors according to multi-dimensional business domains, and establish a twin association graph; An improved denoising diffusion implicit model is constructed. Based on the block coupling module, cross-block fusion denoising and missing completion are performed on the block twin state vector. The graph prior injection module is introduced to perform relevant subspace focusing reconstruction on the twin association graph. The feasible region projection module is used to project the process window and equipment capacity boundary to obtain the twin state reconstruction results and short-term state prediction results. Based on the twin state reconstruction results and short-term state prediction results, an abnormal evidence set is generated, a restart vector is constructed, and the twin association graph is converted into a transition matrix; Based on the restart random walk algorithm, the steady-state equations of the transition matrix and restart vector are directly solved. Core subgraph compression and back-substitution calculation are performed through reverse elimination propagation to obtain the root cause candidate ranking and the set of influencing paths. By ranking root cause candidates and generating a set of impact paths, a cross-process linkage control strategy is generated, issued and executed, and the twin status, twin relationship diagram and abnormal evidence set are updated.

2. The intelligent control method for the entire steel pipe processing process based on digital twins according to claim 1, characterized in that, The production process data includes process parameter data, equipment operation data, quality inspection data, work order batch data, and logistics cycle data.

3. The intelligent control method for the entire steel pipe processing process based on digital twins according to claim 1, characterized in that, The formation of the multi-source fusion dataset includes: Production process data is collected from each step of the entire steel pipe processing process, and batch identifier, process identifier, equipment identifier and timestamp are generated for each data record; The production process data is uniformly encoded and time-aligned. Data with different sampling frequencies are mapped to the same time axis according to the timestamp. Data with multiple records corresponding to the same time axis position are merged, and data with no corresponding record at the time axis position are marked as missing. The time-aligned production process data is preprocessed to form a multi-source fusion dataset. The preprocessing includes unifying units and dimensions, removing outliers and recording removal marks, marking missing fields and recording missing position indexes, and aggregating the data by batch identifier and process identifier to generate the multi-source fusion dataset.

4. The intelligent control method for the entire steel pipe processing process based on digital twins according to claim 1, characterized in that, The construction of the block-based twin state vector and the establishment of the twin association graph include: Based on the multi-source fusion dataset, a block-based twin state vector is generated according to the batch identifier and process identifier. The block-based twin state vector includes process block, equipment block, quality block, work order batch block and logistics cycle block. Each block is formed by concatenating the corresponding fields in dimensional order. Missing fields in each block are represented by the position code corresponding to the missing marker. The process node set is determined based on the process identifier, the equipment node set is determined based on the equipment identifier, the parameter node set is determined based on the process parameter field, the quality node set is determined based on the quality inspection field, the defect node set is determined based on the defect category field, and the batch node set is determined based on the batch identifier. The node sets are then merged to form the node set of the twin association diagram. Based on the multi-source fusion dataset and the block twin state vector, an edge set of twin association graph is generated. The edge set includes process sequence edges, equipment membership edges, parameter effect edges, detection mapping edges, and historical co-occurrence edges.

5. The intelligent control method for the entire steel pipe processing process based on digital twins according to claim 1, characterized in that, The obtained twin state reconstruction results and short-term state prediction results include: An improved denoising diffusion implicit model is constructed, which includes a diffusion process module, a denoising module, a block coupling module, a graph prior injection module, and a feasible region projection module. The diffusion process module generates a forward noise-adding sequence and a reverse noise-de-noising sequence. The forward noise-adding sequence covers time step zero to the end of time step, and generates a noisy state vector for each time step. The noisy state vector is obtained by linear combination of the original state vector and the standard normal noise vector. The coefficient of the original state vector is the square root of the cumulative hold coefficient, and the coefficient of the standard normal noise vector is the square root minus the cumulative hold coefficient. The cumulative hold coefficient is determined by the noise schedule. The reverse noise-de-noising sequence covers time step end to time step zero, and updates the noisy state vector step by step. The block coupling module performs cross-block fusion denoising and missing completion processing on the block twin state vectors. It encodes the process block, equipment block, quality block, work order batch block and logistics cycle block to obtain block representation vectors. Based on the inter-block interaction mapping, the block representation vectors are synthesized into a fused representation vector. The graph prior injection module generates graph prior vectors based on the Siamese association graph, and performs dimension selection operations on the fused representation vectors to obtain focused representation vectors; The denoising module takes the noisy state vector, time step and focused representation vector as input, and outputs the noise estimation vector. Based on the inverse solution of the noise estimation vector, the reconstruction estimation vector of the original state vector is obtained, forming the twin state reconstruction result and the short-term state prediction result. The feasible region projection module performs process window and equipment capacity boundary projection processing on the twin state reconstruction results and short-term state prediction results.

6. The intelligent control method for the entire steel pipe processing process based on digital twins according to claim 1, characterized in that, The construction of the restart vector, which converts the twin association graph into a transition matrix, includes: An abnormal evidence set is generated based on the twin state reconstruction results and short-term state prediction results. The abnormal evidence set is obtained by comparing process parameter values ​​with the upper and lower bounds of the process window, comparing equipment state values ​​with equipment capacity boundaries, comparing quality index values ​​with quality threshold ranges, matching defect category fields, and comparing logistics cycle time values ​​with target cycle time ranges. A restart vector is constructed based on the set of abnormal evidence. The dimension of the restart vector is consistent with the number of nodes in the Siamese graph. The node position corresponding to the abnormal evidence is assigned a value of one, and the remaining node positions are assigned a value of zero. The normalized restart vector is obtained by dividing each dimension value by the sum of the dimension values. The twin graph is converted into a transition matrix. The dimension of the transition matrix is ​​the same as the number of nodes in the twin graph. When there is an edge, the element takes the edge weight; when there is no edge, the element takes zero. The row-normalized transition matrix is ​​obtained by dividing each row element by the sum of the elements in the current row.

7. The intelligent control method for the entire steel pipe processing process based on digital twins according to claim 1, characterized in that, The obtained root cause candidate ranking and influence path set include: Input the transition matrix, restart vector, and restart probability to construct the full node correlation distribution vector and determine the identity matrix with the same dimension as the transition matrix; The steady-state equation is solved directly. The steady-state equation satisfies the condition that the product of the identity matrix minus the product of the restart probability and the transpose of the transition matrix, multiplied by the correlation distribution vector, equals the product of the restart probability and the restart vector. The correlation distribution vector of all nodes is obtained by solving the sparse linear equation system. Core subgraph compression and back-substitution calculation are performed through reverse elimination propagation. Core subgraph compression includes determining the set of eliminateable nodes and the set of retained nodes based on the sequence of processes and the affiliation of equipment. Node elimination is performed on the set of eliminateable nodes to obtain the core subgraph transition matrix and generate the back-substitution mapping relationship. Back-substitution calculation includes calculating the correlation distribution of retained nodes on the core subgraph and restoring the correlation distribution of eliminateable nodes based on the back-substitution mapping relationship. The root cause candidate ranking and the set of influencing paths are generated based on the correlation distribution vector of all nodes. The root cause candidate ranking is obtained by sorting the correlation distribution vectors from largest to smallest value. The set of influencing paths is determined by the set of reachable paths from the node corresponding to the restart vector to the root cause candidate node in the core subgraph transition matrix, and is obtained by sorting the cumulative values ​​of the correlation distribution vectors on the path from largest to smallest value.

8. The intelligent control method for the entire steel pipe processing process based on digital twins according to claim 1, characterized in that, The generation and execution of the cross-process linkage control strategy includes: The root cause candidate node set is determined based on the root cause candidate ranking, and the process node, equipment node, parameter node and quality node set corresponding to the root cause candidate node set is determined based on the influence path set. Cross-process linkage control strategies are generated by combining process nodes, equipment nodes, parameter nodes, and quality nodes. These strategies include process parameter adjustment instructions, equipment operation adjustment instructions, stricter inspection instructions, and process handling instructions. Issue cross-process linkage control strategies and collect execution results. The execution results include execution process identifiers, execution equipment identifiers, execution parameter settings, and execution detection results. Write the execution results into the twin status and update the node status, edge set, and abnormal evidence set of the twin association graph based on the execution results.

9. The intelligent control system for the entire steel pipe processing process based on digital twins according to claim 1, executing the intelligent control method for the entire steel pipe processing process based on digital twins according to any one of claims 1 to 8, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to collect data from the entire steel pipe processing production process and complete preprocessing to form a multi-source fusion dataset. The twin state construction module is used to construct block twin state vectors based on multi-source fusion datasets and establish twin association graphs. The twin reconstruction prediction module is used to construct an improved denoised diffusion implicit model, and generates twin state reconstruction results and prediction results through block twin state vectors and twin association graphs; The Anomaly Evidence and Transition Modeling Module is used to generate anomaly evidence sets and construct restart vectors based on reconstruction and prediction results, and to convert twin correlation graphs into transition matrices. The root cause tracing calculation module is used to complete the direct solution and reverse elimination propagation of the steady-state equation based on the restart random walk algorithm, according to the transition matrix and restart vector, and output the root cause candidate ranking and the set of influence paths; The linkage control execution module is used to generate cross-process linkage control strategies based on the root cause candidate ranking and the set of impact paths, and to issue and execute them, and update the twin status, twin relationship diagram and abnormal evidence set.

Images