A steel structure manufacturing deviation traceability and self-optimization method and system based on multi-source data fusion

Abstract

The application discloses a kind of steel structure manufacturing deviation traceability and self-optimization method and system based on multi-source data fusion, including the multi-modal data of multi-source acquisition steel structure manufacturing process, digital twin model is constructed based on process mechanism and real-time data;Through the fusion analysis and traceability reasoning of multi-modal data by explainable artificial intelligence engine, deviation traceability vector with evidence chain is generated, and digital twin model is driven to map process state in real time;Through the deviation propagation simulation deduction and multi-link root cause decoupling of traceability vector input twin model, the explainable diagnosis conclusion and accurate process optimization strategy are generated;Based on optimization strategy, control instruction is dynamically generated by model factory mechanism, parameter self-adjustment and process self-optimization are carried out;The present application solves the fundamental defects of existing system in AI decision black box, inefficient data fusion, single-point optimization limitation and industry knowledge island through the closed-loop architecture of perception, traceability, optimization and evolution.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and industrial digital twin technology, specifically involving a method and system for tracing and self-optimizing deviations in steel structure manufacturing based on multi-source data fusion. Background Technology

[0002] Steel structures are the core load-bearing system for large buildings, bridges, and industrial facilities. Their manufacturing and assembly precision directly determines the safety, durability, and functionality of the overall structure. The manufacturing process of steel structures involves many complex procedures such as material cutting, assembly, welding, and straightening. Deviations in each stage can be transmitted, coupled, and amplified along the manufacturing chain, ultimately leading to serious quality problems such as geometric deviations of components, residual stress concentration, and decreased connection performance. As modern engineering structures become larger and more complex, the requirements for the manufacturing precision and consistency of steel components are becoming increasingly stringent. Traditional quality control models that rely on manual experience and post-inspection are no longer sufficient to meet these demands. Existing quality control technologies for steel structure manufacturing processes have significant limitations in terms of comprehensive perception, precise traceability, and proactive control of deviations. At the data perception level, various sensors deployed on the manufacturing site, such as vision, force, and temperature sensors, typically operate independently, generating multimodal data, including 3D point clouds, process parameter time series, and physical field distributions, which are heterogeneous in terms of spatiotemporal reference, format, and semantics. This lack of an effective fusion mechanism makes it difficult to construct a high-fidelity digital mapping that reflects all elements and states of the manufacturing process. This leads to data silos, failing to provide a complete and consistent data foundation for deviation analysis. At the level of deviation analysis and source tracing, existing methods mostly rely on statistical process control or simulation analysis of a single physical field, lacking interpretable intelligent analysis capabilities that integrate multi-source data, process mechanisms and data-driven models. When faced with manufacturing deviations, it is often difficult to quickly and accurately locate the process links and root cause parameters that cause them, and it is even more difficult to quantify the contribution of multiple factors coupled together, making corrective measures trial-and-error in nature, inefficient and with unstable effects. At the execution and optimization level, existing control systems typically adjust parameters based on fixed process procedures or simple feedback, lacking decision support based on digital twin simulation and forward-looking extrapolation. The system cannot simulate, verify, and evaluate candidate optimization strategies in the digital space, nor can it achieve closed-loop autonomous optimization from deviation diagnosis to process parameter self-adjustment, resulting in insufficient adaptability, flexibility, and intelligence in the manufacturing process. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for tracing and self-optimizing manufacturing deviations in steel structures based on multi-source data fusion. By constructing a digital twin model that is synchronized with the physical manufacturing process in real time, and integrating interpretable artificial intelligence and forward-looking simulation and deduction technology, the invention achieves automatic and accurate tracing of manufacturing deviations and adaptive closed-loop optimization of process parameters. This effectively solves the technical problems in traditional manufacturing quality control, such as fragmented multi-source data, difficulty in locating the root cause of deviations, and reliance on trial and error and experience for process adjustments. As a result, the invention significantly improves the accuracy, consistency, and intelligence level of steel structure manufacturing.

[0004] The objective of this invention can be achieved through the following technical solutions: This application provides a method for tracing and self-optimizing manufacturing deviations in steel structures based on multi-source data fusion, including the following steps: Multi-modal data from multiple sources are collected during the steel structure manufacturing process, and a digital twin model of the steel structure manufacturing process is constructed and updated synchronously based on the preset process mechanism and the multi-modal data. The multimodal data includes three-dimensional point cloud and visual image data, stress-strain and temperature field data, process timing data and environmental condition data; Based on the digital twin model, the multimodal data is fused, analyzed, and traced through an interpretable artificial intelligence engine to generate a deviation traceability vector with a physical evidence chain. Based on the deviation state represented by the traceability vector, it is dynamically mapped to the corresponding parameter space of the digital twin model. Simultaneously, the deviation source vector is input into the digital twin model that has completed dynamic mapping to perform simulation deduction of deviation propagation and multi-stage root cause decoupling analysis, generating interpretable diagnostic conclusions and corresponding process optimization strategies. Based on the aforementioned process optimization strategy, adaptive control commands are dynamically generated through a model factory mechanism to perform online self-adjustment of the manufacturing process parameters of the manufacturing system. Among them, the manufacturing process status data after the manufacturing process parameters are adjusted is used as new multimodal data input into the digital twin model to verify and iteratively optimize the accuracy of the digital twin model and the process optimization strategy; Furthermore, based on the preset process mechanism, an initial digital twin model including a geometric model, a physical model, and a behavioral model is constructed; during the steel structure manufacturing process, the initial digital twin model is synchronously updated using real-time acquired multimodal data, including: based on real-time acquired 3D point cloud and visual image data, registration and comparison are performed with the geometric model using a point cloud registration algorithm to generate a geometric deviation field, and the geometric model is corrected in real time in terms of shape and size based on the geometric deviation field; Furthermore, based on the real-time collected stress, strain, and temperature field data, the physical model is used as the input of real-time loads and boundary conditions. Physical simulation is performed through a finite element solver to drive the real-time updating of the internal mechanical and thermal states of the physical model. Based on the real-time collected process timing data and environmental state data, the process logic state and performance evolution path of the behavior model are driven to be updated in real-time through rule reasoning. Furthermore, feature point sets representing key geometric shapes of steel structure components are extracted from 3D point cloud and visual image data, and corresponding design reference models are obtained from the geometric models; the feature point sets and the design reference models are matched based on feature descriptors to complete the initial coarse registration. Simultaneously, using the coarse registration result as the initial pose, the 3D point cloud, visual image data, and design reference model are iteratively registered to the nearest point to solve for the optimal spatial transformation matrix. The optimal spatial transformation matrix is ​​then applied to calculate the positional and normal deviations between the 3D point cloud, visual image data, and design reference model, and these deviations are fused to generate a structured geometric deviation field with a continuous spatial distribution. Furthermore, the real-time collected stress, strain, and temperature field data are associated with the geometric region corresponding to the physical model to generate the real-time boundary conditions and physical loads required for the finite element simulation model. Based on the real-time boundary conditions and physical loads, the finite element solver performs a thermo-mechanical sequential coupling solution on the finite element simulation model, sequentially performing transient heat conduction analysis and thermo-mechanical analysis to obtain the transient temperature field, stress field, and strain field during the manufacturing process. The transient temperature field, stress field, and strain field are synchronized to the physical model as real-time internal states to drive the updating of the mechanical and thermal fields within the physical model. Furthermore, through the rule parsing engine, the real-time collected process timing data is parsed into specific process instructions and status events, and the environmental status data is matched with the preset environmental tolerance threshold in the behavior model to generate environmental trigger events. The process instructions, state events, and environmental trigger events are input into the behavior model. Based on the process rules preset in the behavior model, the input events are matched with the state transition conditions in the preset process rules to determine the current process logic state and deduce the subsequent performance evolution path. According to the deduction results, the current state node and evolution path parameters of the behavior model are updated to complete the real-time logical synchronization with the physical manufacturing process. Furthermore, the multimodal data is subjected to spatiotemporal alignment and standardization, and key features are extracted from the multimodal data to construct a unified multidimensional feature vector; the key features include at least: deformation mode statistics extracted from the geometric deviation field, thermo-mechanical coupling feature indices extracted from the stress-strain and temperature field data, and time-series event features extracted from the process time series and environmental state data; Furthermore, the multidimensional feature vectors are input in parallel into the data-driven prediction model and the process knowledge graph in the interpretable artificial intelligence engine; wherein, the data-driven prediction model is input to obtain preliminary hypotheses about the causes of deviations and their confidence levels; and the process knowledge graph is input simultaneously to generate candidate causal chains through graph traversal and logical reasoning. The preliminary hypothesis of the cause of the deviation is verified and fused with the candidate causal chain to generate the deviation source vector, which includes the root cause process identifier, the associated physical process parameters and their contribution, and an attached physical evidence chain. The verification and fusion process includes at least: sorting the deviation cause hypotheses according to their confidence levels, and verifying the consistency between the deviation cause hypotheses and the causal chain according to the logical relationships in the process knowledge graph; wherein, the physical evidence chain is associated with the multimodal data, deformation mode statistics, thermo-mechanical coupling characteristic indicators, time-series event characteristics, and inference paths in the process knowledge graph corresponding to the deviation source vector through reference relationships. Furthermore, based on the deviation tracing vector, the physical process parameters identified by the vector that cause the steel structure manufacturing deviation are located and adjusted in the digital twin model, so as to inject the initial deviation into the corresponding manufacturing process in the digital twin model. Starting from the point where the deviation is injected, the digital twin model performs a multiphysics time-series simulation of the steel structure manufacturing process to deduce the propagation and evolution of the initial deviation along the actual process chain to subsequent stages, and obtains the geometric and mechanical deviations output by each stage during the simulation as the basic data for decoupling analysis. Based on the aforementioned basic data, a contribution decoupling analysis is performed on multiple suspected root causes identified in the deviation tracing vector. Through quantitative calculation, the independent influence and interactive effect of each suspected root cause under the coupling effect of multiple manufacturing links are distinguished, and a process optimization strategy that clearly points to specific process links and parameter adjustment directions is generated. Furthermore, by receiving and parsing the process optimization strategy, control semantic information regarding the target process, the parameters to be adjusted, and the adjustment target is extracted. Based on the control semantic information, the pre-set process control logic model in the model factory mechanism is invoked to generate a specific control logic sequence that matches the current manufacturing equipment and operating conditions. The control logic sequence is mapped to the control interface protocol of the target device and encapsulated into a standardized instruction package, which is then transformed into adaptive control instructions that the manufacturing system can recognize and execute. The adaptive control instructions include at least real-time parameter settings and action instructions for welding equipment, assembly tooling and straightening mechanism. As a preferred embodiment, a steel structure manufacturing deviation tracing and self-optimization system based on multi-source data fusion includes: The multi-source data acquisition module is used to acquire multimodal data during the steel structure manufacturing process. The multimodal data includes three-dimensional point cloud and visual image data, stress-strain and temperature field data, process timing data and environmental condition data. The digital twin model construction and synchronization module is used to construct and synchronously update a digital twin model of the steel structure manufacturing process based on the preset process mechanism and the multimodal data. An interpretable artificial intelligence engine module is used to perform fusion analysis and source tracing reasoning on the multimodal data, generate a deviation source tracing vector with a physical evidence chain, and dynamically map the deviation state represented by the source tracing vector to the corresponding parameter space of the digital twin model. The simulation and strategy generation module is used to input the deviation source vector into the digital twin model that has completed dynamic mapping, perform simulation and decoupling analysis of deviation propagation and multi-stage root cause analysis, and generate interpretable diagnostic conclusions and corresponding process optimization strategies. The model factory control module is used to dynamically generate adaptive control commands based on the process optimization strategy through the model factory mechanism, and to perform online self-adjustment of the manufacturing process parameters of the manufacturing system. The closed-loop verification and iterative optimization module is used to input the manufacturing process status data after adjusting the manufacturing process parameters as new multimodal data into the digital twin model, and to verify and iteratively optimize the accuracy of the digital twin model and the process optimization strategy.

[0005] The beneficial effects of this invention are as follows: This invention achieves high-fidelity dynamic mapping and perception of the entire manufacturing process by constructing a real-time synchronized digital twin model and integrating multi-source heterogeneous data, effectively solving the data silo problem. Furthermore, by integrating a data-driven model and an interpretable artificial intelligence engine with a process knowledge graph, it can generate deviation source vectors with attached physical evidence chains and quantified contribution, achieving accurate and reliable location of the root causes of manufacturing deviations and overcoming the shortcomings of traditional methods that rely on experience and are difficult to attribute. Based on simulation and decoupling analysis of multiple stages using a digital twin model, this invention can predict deviation propagation in advance and generate validated optimization strategies. Through the model factory mechanism, the strategies are automatically converted into adaptive control commands, driving the manufacturing system to adjust online, forming a rapid closed loop of perception, analysis, decision-making, and execution. This significantly improves the scientific nature and predictability of process adjustments, as well as the autonomous optimization and fault tolerance capabilities of the manufacturing system. Attached Figure Description

[0006] To better understand and implement this application, the technical solution is described in detail below with reference to the accompanying drawings.

[0007] Figure 1 A flowchart illustrating a method for tracing and self-optimizing manufacturing deviations in steel structures based on multi-source data fusion, provided for this application; Figure 2 A flowchart illustrating the process of generating and updating a digital twin model based on a multi-source data fusion-based method for tracing and self-optimizing steel structure manufacturing deviations, provided in this application. Figure 3 This is a flowchart illustrating a steel structure manufacturing deviation tracing and self-optimization system based on multi-source data fusion. Detailed Implementation

[0008] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description relating to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of methods and systems consistent with some aspects of this application as detailed in the appended claims.

[0009] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0010] The following detailed description of the specific embodiments, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0011] Example 1 Please see Figures 1-3This embodiment details a method for tracing and self-optimizing manufacturing deviations in steel structures based on multi-source data fusion. By constructing a real-time synchronized digital twin model and integrating interpretable artificial intelligence analysis, it achieves automatic tracing of manufacturing deviations and process self-optimization.

[0012] S1. Collect multi-modal data from multiple sources during the steel structure manufacturing process, and construct and synchronously update a digital twin model of the steel structure manufacturing process based on the preset process mechanism and the multi-modal data. S1.1 Multimodal Data Acquisition: A distributed sensing system is deployed on the steel component welding and assembly production line to synchronously acquire the multimodal data, specifically including: 3D point cloud and visual image data: Install a high-precision structured light 3D scanner (such as GOM ATOS Q) at key inspection stations (such as after assembly and welding) to collect dense 3D point clouds on the surface of components at a frequency of 2Hz. The number of point clouds in a single frame is no less than 2 million points, while triggering an industrial camera to collect high-resolution color images. Stress, strain and temperature field data: A network of resistance strain gauges was attached near the weld and in the heat-affected zone, supplemented by a distributed fiber optic grating sensor, to simultaneously measure dynamic strain and temperature at a sampling rate of 1000 Hz; an infrared thermal imager was used to monitor the two-dimensional temperature field distribution of the weld pool and surrounding area. Process timing data: Real-time time series data of key process parameters such as welding current (I), arc voltage (U), welding speed (v), and clamping force (F) are read from the welding power source, robot controller, and tooling fixture PLC via industrial Ethernet and OPC UA protocol, with a sampling frequency of not less than 500Hz; Environmental status data: A wireless sensor node network is deployed in the workshop to monitor ambient temperature (T_env), relative humidity (RH), and ground vibration acceleration (a). The data is transmitted back through a LoRa gateway. All data streams are timestamped to the millisecond level by a unified NTP time server and bound to a unique ID of the component identified by RFID, ensuring spatiotemporal alignment of multi-source data. S1.2 The steps for building and synchronously updating the digital twin model are as follows: Constructing an initial digital twin model: Based on the CAD design model, material property library, and process documents, construct an initial digital twin model including a geometric model, a physical model, and a behavioral model. Geometric model: A design CAD model in B-Rep format; Physical model: A thermo-mechanical coupling simulation model based on the finite element method, including material constitutive relations, welding heat source model and contact conditions; Behavioral Model: An extended finite state machine is used to define the manufacturing process flow, process logic, and state transition rules; During the manufacturing process, the initial digital twin model is updated synchronously in real time. The specific steps are as follows: Geometric model update: Driven by real-time acquired 3D point cloud and visual image data; Feature extraction and coarse registration: ISS feature points are extracted from 3D point cloud and visual image data, and corresponding point sets are generated by sampling from the surface of CAD model. FPFH descriptors are calculated for matching to complete the initial coarse registration and obtain the initial transformation matrix T_init. Fine registration and bias field generation: Starting from T_init, fine registration is performed using the iterative nearest-point algorithm, and the optimal transformation matrix T_opt is solved by minimizing the objective function; the objective function is specifically: Where p_i is the measured 3D cloud point, q_i is the corresponding nearest point on the CAD model, and R and t are the rotation matrix and translation vector to be determined; after applying T_opt, the Euclidean distance deviation d_i and normal deviation θ_i from each point cloud point to the surface of the CAD model are calculated, and a structured geometric deviation field D(x,y,z) = {d,θ} is generated through spatial interpolation; Model correction: Based on the geometric deviation field D, the geometric model is deformed and rendered to intuitively display the actual geometric state; Physical model update: Driven by real-time acquired stress, strain and temperature field data; Data association and boundary condition generation: The positions of strain gauges and FBG sensors are mapped to the finite element mesh nodes of the physical model, and their measured values ​​(strain ε_meas, temperature T_meas) are associated as the real-time boundary conditions of the corresponding nodes. Physical simulation solution: The integrated finite element solver is invoked to perform a thermo-mechanical sequential coupling analysis; first, a transient thermal analysis is performed to solve for the temperature field T(x,y,z,t) within the domain Ω, with the governing equations as follows: Where ρ is density, c_p is specific heat capacity, k is thermal conductivity, and Q is internal heat source; then the obtained temperature field is used as a thermal load for mechanical analysis to solve for stress field σ and strain field ε. State synchronization: The transient temperature field, stress field and strain field results obtained from the simulation calculation are updated in real time to the visualization module of the digital twin physical model; Behavioral model update: Driven by process timing data and environmental status data; Event parsing: The rule parsing engine parses the time-series data stream into discrete process events (such as welding start, current over-limit), and compares environmental data with thresholds to generate environmental events (such as humidity warning). State deduction: Input the above events into the behavior model and match them with predefined process rules (such as triggering the start of the next welding state transition if an interlayer temperature reaches the standard event) to determine the current process logic state and deduce possible subsequent evolution paths (such as predicting the remaining time of the current process). Logical synchronization: Based on the deduction results, update the active state, timer and performance parameters of the behavioral model to achieve logical synchronization with the physical process; Furthermore, through step S1, by acquiring multi-source data and constructing and synchronizing the digital twin model in real time, this method constructs a digital mainline covering the entire manufacturing process: by deploying a distributed sensor network and unifying the spatiotemporal reference, it solves the problem of data silos formed by the isolation of multimodal data in terms of source, format, and time, and realizes the synchronous acquisition and correlation of all elements of three-dimensional geometry, physical field, process sequence, and environmental state; then, based on the preset process mechanism and the real-time inflow of multi-source data, it drives the continuous updating of the digital twin model containing geometric, physical, and behavioral models, solving the problems of static, lagging, and disconnected from the physical process of traditional models, realizing high-fidelity dynamic mapping and real-time full-state visualization of the manufacturing process from physical space to digital space, and providing an accurate, consistent, and vivid data foundation and model environment for subsequent intelligent analysis, simulation, and closed-loop control; S2. Based on the digital twin model, the multimodal data is fused, analyzed and reasoned through an interpretable artificial intelligence engine to generate a deviation source vector with a physical evidence chain, and the deviation state represented by the source vector is dynamically mapped to the corresponding parameter space of the digital twin model. The specific implementation steps for using an interpretable artificial intelligence engine to perform fusion analysis and source tracing reasoning on the multimodal data to generate a bias source tracing vector with an attached physical evidence chain are as follows: Feature extraction and construction of multidimensional feature vectors: Extract key features from the aligned and synchronized multimodal data in S1 and construct a unified feature vector F; The key features include: deformation mode statistics extracted from the geometric deviation field D, thermo-mechanical coupling feature indices extracted from physical field data, and time-series event features extracted from time-series and environmental data; The deformation mode statistics include: overall deformation energy, local maximum deviation, and curvature change statistics; The thermo-mechanical coupling characteristic indicators include: maximum residual principal stress, mean transverse temperature gradient, and key points of cooling rate. The time-series event characteristics include: the time-series fluctuation characteristics of welding heat input (η=U*I / v), the time difference of key processes, and the average ambient temperature and humidity; The feature vector F is then input in parallel to two core components of the interpretable artificial intelligence engine; the core components include: a data-driven prediction model and a process knowledge graph; Specifically, the explainable AI engine refers to a hybrid reasoning framework that integrates a data-driven prediction model and a process knowledge graph. Explainability is achieved through the following collaborative mechanism: the data-driven prediction model (in this embodiment, a gradient boosting tree) learns complex patterns from historical data and outputs preliminary hypotheses about the causes of deviations and the confidence level of each hypothesis; the process knowledge graph is constructed based on domain expert knowledge, using a graphical structure to solidify the causal relationship rules between processes, parameters, and physical quantities; the explainable AI engine combines the possibilities of data-driven reasoning with the causal logic of the knowledge graph through parallel reasoning and result verification, ultimately generating conclusions with accompanying evidence chains, thereby solving the problems of pure data model black boxes and rigid pure knowledge reasoning. Furthermore, the feature vector F is input data to drive the prediction model (a trained gradient boosting tree model), and outputs a preliminary set of hypotheses about the root causes of bias, H_data = {(cause_i, confidence_i)}. Meanwhile, the feature vector F is input into the process knowledge graph, which contains the causal and correlation relationships between processes, parameters, physical quantities and defects; the AI ​​engine can traverse and logically reason in the process knowledge graph based on the feature vector F to generate a set of candidate causal chains H_kg that conforms to the mechanism. Verification and Fusion: H_data and H_kg are verified and fused. For example, if the data-driven prediction model gives a high-confidence hypothesis about abnormal wire feed speed, but process knowledge graph inference finds that the wire feed speed is not a sensitive parameter in the current welding stage, then the hypothesis about abnormal wire feed speed is downweighted. The fusion process, based on confidence ranking and logical consistency verification, ultimately generates a structured deviation tracing vector V. Where: P_root: root cause process; { (p_j, v_j, c_j)}: a set of triples, each triple representing the name, current value, and contribution of a relevant process parameter; E_chain: physical evidence chain; The physical evidence chain E_chain is linked to the multimodal data, deformation mode statistics, thermo-mechanical coupling characteristic indicators, time-series event characteristics, and reasoning paths in the process knowledge graph corresponding to the generation of the deviation tracing vector through reference relationships. Finally, based on the root cause process P_root identified in the source vector V and the associated parameter list { (p_j, v_j, c_j)}, the corresponding physical process parameters are located in the digital twin model (for example, the zero-position offset parameter of the welding robot's second axis is located in the behavior model of the welding process). Subsequently, the system sets the values ​​of these parameters in the model to the actual values ​​recorded in vector V that cause the deviation, so as to accurately reproduce the parameter state that causes the deviation in the digital space, thereby completing the dynamic mapping between the model parameter space and the physical deviation state, and providing accurate input conditions for subsequent simulation and deduction. Furthermore, in step S2, through fusion analysis and dynamic mapping based on an interpretable artificial intelligence engine, and by integrating data-driven approaches and knowledge mechanisms, the problem of the disconnect between black-box prediction and causal logic in manufacturing deviation diagnosis is solved: by inputting a unified feature vector extracted from multimodal data into the data-driven prediction model and the process knowledge graph in parallel, the problems of poor interpretability of a single data model and the inability of pure knowledge reasoning to cope with complex nonlinear relationships are solved; by verifying and fusing the outputs of the two based on confidence and logical consistency, a deviation source vector with physical evidence chain and quantified contribution is generated, solving the problem that traditional methods are difficult to accurately locate the root cause process and parameters in complex multi-factor coupling and that the conclusions lack traceable physical basis, thus achieving accurate, interpretable, and quantifiable attribution of the root cause of deviation; and by dynamically mapping this vector to the corresponding parameter space of the digital twin model, the physical deviation source is accurately reproduced in the digital space, establishing high-fidelity initial conditions for subsequent simulation and deduction, and completing the intelligent leap from data and phenomena to models and causes. S3. Simultaneously, the deviation source vector is input into the digital twin model that has completed dynamic mapping to perform simulation deduction of deviation propagation and multi-stage root cause decoupling analysis, generating interpretable diagnostic conclusions and corresponding process optimization strategies. Specifically, the implementation steps for the simulation and decoupling analysis of deviation propagation and multi-stage root cause analysis, based on the input of the deviation source vector into the dynamically mapped digital twin model, are as follows: Based on the deviation source vector V, the root cause process (such as cutting and blanking) is located in the digital twin model, and the corresponding physical process parameters are adjusted. Starting from the root cause process, rerun the multiphysics time-series simulation covering all subsequent related processes (such as assembly and welding); simulate how the initial deviation (such as the deviation of the cut surface angle caused by improper cutting speed) propagates and evolves along the process chain, and record the geometric deviation δ_geom and mechanical deviation δ_mech output at each link as basic data; Multi-stage root cause decoupling analysis: Using the basic data obtained from simulation, multiple possible root causes in V are decoupled; by designing simulation experiments, the independent influence Δ_single of changes in a single factor (such as changing only the cutting speed) on the final deviation is quantified. Full factorial simulation was used to analyze the total deviation under the combined effects of multiple factors, and the interaction effect Δ_interaction was calculated. Based on independent effects and interaction effects, the contribution of each root cause was calculated. Based on the aforementioned contribution decoupling analysis, interpretable diagnostic conclusions (identifying primary and secondary causes and quantifying contributions) and corresponding process optimization strategies are generated. Furthermore, in step S3, through simulation, root cause decoupling, and strategy generation, and by conducting forward-looking experiments in the digital twin model, the decision-making bottleneck from deviation attribution to the formulation of reliable optimization schemes is solved: by locating the deviation source vector generated in S2 in the digital twin model and injecting the corresponding deviation, the problems of high cost, long cycle, and difficulty in reproducing complex working conditions of physical experiments are solved, realizing high-fidelity digital reproduction and active experimentation of the deviation generation and propagation process; then, multi-physics time-series simulation covering the entire process chain is executed to deduce the dynamic propagation and coupling evolution of the initial deviation in multiple manufacturing links, solving the limitation of traditional methods that can only analyze isolated links and are difficult to predict the chain effect of deviations, and realizing quantitative prediction of the propagation path and consequences of the entire manufacturing deviation; based on simulation data, contribution decoupling analysis of multiple suspected root causes is performed, and the independent influence and interaction effect of each factor are distinguished through quantitative calculation, solving the problem of fuzzy root cause contribution and difficulty in determining adjustment priority under multi-factor coupling, and finally generating a process optimization strategy that is clearly targeted, quantitatively reliable, and digitally verified; S4. Based on the process optimization strategy, adaptive control instructions are dynamically generated through the model factory mechanism to perform online self-adjustment of the manufacturing process parameters of the manufacturing system; Specifically, the model factory mechanism is a software adaptation layer that automatically compiles high-level, abstract process optimization strategies into executable instructions for low-level equipment; specifically, it is a repository of pre-built process control logic models, each corresponding to a specific type of equipment (such as a certain type of welding robot) or process scenario (such as flat welding). Each process control logic model encapsulates the control interface protocol, parameter mapping relationship, and basic control algorithm (such as PID parameter self-tuning logic) of the specific equipment. Furthermore, the implementation process of dynamically generating adaptive control instructions through the model factory mechanism is as follows: The model factory mechanism receives the process optimization strategy generated by S3 and extracts the control semantics, such as: target process: welding; equipment to adjust: robot #1; parameter to adjust: welding current; target value: I_target; Based on the control semantics, the system calls the pre-set process control logic model (such as a welding parameter control model containing PID adjustment logic) in the model factory that matches the robot #1 model to generate a specific control logic sequence. This control logic sequence is encapsulated into a standardized instruction package through the device protocol mapping layer; the standardized instruction package is then sent to the target device (welding robot) via the industrial network to adjust its manufacturing process parameters in real time (setting the welding current to I_target), thus completing online self-adjustment; Furthermore, step S4, through adaptive control instruction generation and online self-adjustment based on the model factory mechanism, and by constructing an intelligent compilation link from strategy to execution, solves the execution bottleneck of disconnect between optimization decision-making and actual control, and system response lag. It receives and parses the simulation-verified process optimization strategy generated in S3, extracting semantic information about the target process, parameters to be adjusted, and adjustment targets, thus resolving the semantic gap problem where high-level optimization targets cannot be directly understood and executed by the underlying manufacturing equipment. Based on this, it calls the pre-set process control logic model in the model factory, which matches the specific equipment and operating conditions, to dynamically... The system generates specific control logic sequences, overcoming the limitations of traditional control systems that rely on fixed and universal control programs and lack flexible adaptability. Furthermore, the system encapsulates these logic sequences into standardized adaptive control instructions through a device protocol mapping layer and distributes them to execution units such as welding equipment, assembly fixtures, or straightening mechanisms. This solves the integration problem of inconsistent control interfaces and heterogeneous instructions for multi-source heterogeneous equipment, ultimately achieving seamless, accurate, and rapid transformation from digital optimization strategies to physical execution actions. This completes the closed loop of the decision-making-execution process, enabling the manufacturing system to have online self-adjustment capabilities based on real-time diagnostics and forward-looking simulation. S5. The manufacturing process status data after the manufacturing process parameters are adjusted is used as new multimodal data input to the digital twin model to verify and iteratively optimize the accuracy of the digital twin model and the process optimization strategy. In this process, after the online self-adjustment in step S4, the state data generated in the new round of manufacturing process is collected and input into the digital twin model as new multimodal data; forward simulation is performed through the digital twin model, and the simulation prediction results are compared with the corresponding actual manufacturing results to generate verification results that include strategy effectiveness measurement and model deviation measurement. The effectiveness measure of the strategy: Calculate the deviation reduction rate. This is used to evaluate the actual effectiveness of process optimization strategies; The model bias metric is calculated as a moving average of the prediction error. It is used to evaluate the prediction accuracy of digital twin models; Based on strategy effectiveness measurement The weight parameters of the data-driven prediction model in the interpretable AI engine are updated using reinforcement learning methods (such as policy gradient). Based on model deviation measurement The confidence weights of relevant causal rules in the process knowledge graph are adjusted using a Bayesian update method. Based on the comprehensive verification results, the decision tree induction method is used to optimize the logical rules in the model factory mechanism that map process optimization strategies to control commands; Furthermore, by constructing a learning loop encompassing decision-making, execution, and feedback, the fundamental problems of static solidification and inability to adapt to dynamic changes in the manufacturing process—namely, the inability of digital twin models, analysis engines, and control systems to adapt to dynamic changes in the manufacturing process—are addressed. By using the new round of manufacturing data generated after executing adaptive control commands as feedback input to drive the updating of the digital twin model and forward simulation, and by systematically comparing the predicted results with the actual results, the problems of objectively evaluating the effectiveness of optimization strategies and guaranteeing long-term prediction accuracy of the model are solved. This achieves objective quantitative verification of strategy effectiveness (e.g., deviation reduction rate) and model fidelity (e.g., prediction error). Based on the verification results, reinforcement learning, Bayesian updating, and decision tree induction methods are used to selectively update the data-driven prediction model, the causal rules of the process knowledge graph, and the strategy-command mapping logic. This overcomes the limitations of traditional systems, such as reliance on manual parameter tuning, lagging knowledge base updates, and rigid control logic. Ultimately, this drives collaborative self-learning and continuous performance evolution of the perception, analysis, decision-making, and execution subsystems under real feedback, enabling the entire manufacturing system to possess endogenous intelligence that continuously improves itself from historical experience and real-time interaction, achieving a leap from single-loop control to continuous autonomous optimization. This embodiment constructs an autonomous loop of perception, tracing, decision-making, execution, and optimization: Through multi-source data fusion and real-time synchronization with digital twins, the problem of data fragmentation is solved, achieving high-fidelity real-time perception throughout the manufacturing process; by integrating data and knowledge through an interpretable artificial intelligence engine, the challenge of ambiguous root cause localization of deviations is overcome, achieving accurate and reliable tracing and quantitative attribution; simulation and decoupling analysis based on the digital twin model overcomes the limitations of trial-and-error in process adjustments, generating optimized strategies validated by prior knowledge; adaptive control commands are generated using a model factory mechanism, enabling online and precise self-adjustment of process parameters; closed-loop verification and iterative optimization drive the continuous learning and evolution of the entire system; thus, a series of bottlenecks from perception and diagnosis to control are systematically solved, ultimately achieving a comprehensive improvement and autonomous optimization of the manufacturing process in terms of accuracy, consistency, responsiveness, and intelligence.

[0013] Example 2 Please see Figures 1-2This embodiment is a steel structure manufacturing deviation tracing and self-optimization system based on multi-source data fusion, applied to the steel structure manufacturing deviation tracing and self-optimization method based on multi-source data fusion as described in any one of claims 1-9, specifically including: The multi-source data acquisition module is used to deploy a distributed sensor network to collect and integrate multimodal data from the steel structure manufacturing process, including 3D point cloud and visual image data, stress-strain and temperature field data, process timing data, and environmental condition data; specifically, it is used to perform the following operations: Three-dimensional point cloud and RGB images of component surfaces are acquired using a high-precision laser scanner and industrial camera array; dynamic strain and two-dimensional temperature field distribution in key areas are acquired using a fiber optic sensor network and infrared thermal imager; real-time timing data of process parameters such as welding current, voltage, and speed are read from the equipment control system via the OPC UA protocol; environmental temperature, humidity, and vibration data in the workshop are acquired via a wireless sensor node network; all data are aligned and bound using a unified spatiotemporal reference. The digital twin model construction and synchronization module is used to construct and synchronously update a digital twin model of the steel structure manufacturing process based on the preset process mechanism and the multimodal data; specifically, it is used to perform the following operations: An initial digital twin model is constructed, comprising a geometric model (B-Rep CAD), a physical model (thermo-mechanical coupled finite element model), and a behavioral model (extended finite state machine). Real-time 3D point cloud data is used for registration via iterative nearest neighbor (ICP) algorithm and feature descriptor matching to generate a geometric deviation field, which is then used to correct the shape and size of the geometric model in real time. Real-time acquired stress-strain and temperature field data are input as boundary conditions into the physical model, and a thermo-mechanical sequential coupling simulation is performed using a finite element solver to drive real-time updates of the internal mechanical and thermal states of the physical model. Process timing and environmental state data are parsed into process events, and rule-based reasoning drives the real-time deduction and updating of the process logic state and performance evolution path of the behavioral model, achieving synchronization with the physical manufacturing process. The interpretable artificial intelligence engine module is used to perform fusion analysis and source tracing reasoning on the multimodal data, generate a deviation source tracing vector with an attached physical evidence chain, and dynamically map the deviation state represented by the source tracing vector to the corresponding parameter space of the digital twin model; specifically, it is used to perform the following operations: Multimodal data is spatiotemporally aligned and standardized to extract key features and construct a unified multidimensional feature vector F. F is then input into the data in parallel to drive a prediction model (gradient boosting tree) and a process knowledge graph. A preliminary set of deviation cause hypotheses H_data is obtained through the data-driven model, and a candidate causal chain set H_kg is generated through knowledge graph traversal and logical reasoning. H_data and H_kg are verified and fused based on confidence ranking and logical consistency to generate a deviation tracing vector V containing root cause process identifiers, associated process parameters, and their contributions. The physical evidence chain is linked to the original data, feature indicators, and the graph reasoning path. Based on vector V, the corresponding physical process parameters are located and adjusted in the digital twin model. The simulation and strategy generation module is used to input the deviation source vector into the dynamically mapped digital twin model, perform simulation and decoupling analysis of deviation propagation and multi-stage root cause analysis, and generate interpretable diagnostic conclusions and corresponding process optimization strategies; specifically, it is used to perform the following operations: Based on the deviation source vector V, an initial deviation identified by the vector is injected into the corresponding link of the digital twin model. Starting from this link, a multiphysics time-series simulation of the complete manufacturing process chain is performed to deduce the propagation and evolution of the initial deviation along subsequent links, and the geometric and mechanical deviations output by each link are recorded as basic data. Based on the basic data, the independent influence and interaction effects of multiple suspected root causes identified in vector V are decoupled and analyzed through the control variable method and contribution analysis, and the contribution of each root cause is quantified. An interpretable diagnostic conclusion and process optimization strategy containing clear primary and secondary causes, quantified contribution, and specific parameter adjustment directions are generated.

[0014] The model factory control module is used to dynamically generate adaptive control commands based on the process optimization strategy through the model factory mechanism, and to perform online self-adjustment of the manufacturing process parameters of the manufacturing system; specifically, it is used to perform the following operations: The received process optimization strategy is parsed, and control semantic information about the target process, parameters to be adjusted, and adjustment targets is extracted. Based on the control semantics, a pre-set process control logic model matching the target manufacturing equipment in the model factory is invoked to generate a specific control logic sequence. This control logic sequence is encapsulated into a standardized instruction package through the equipment protocol mapping layer and transformed into an adaptive control instruction that the manufacturing system can recognize and execute. The instruction includes real-time parameter settings and action instructions for welding equipment, assembly fixtures, or straightening mechanisms, and is sent to the corresponding equipment for execution. The closed-loop verification and iterative optimization module is used to input the manufacturing process state data after adjusting the manufacturing process parameters as new multimodal data into the digital twin model, verify and iteratively optimize the accuracy of the digital twin model and the process optimization strategy; specifically, it is used to perform the following operations: The newly generated manufacturing process state data after executing adaptive control commands is used as a new round of multimodal data input to drive the digital twin model update and perform forward simulation. The simulation prediction results are compared with the corresponding actual manufacturing results to generate verification results that include strategy effectiveness metrics (such as deviation reduction rate) and model deviation metrics (such as prediction error). Based on the strategy effectiveness metrics, the weight parameters of the data-driven prediction model in the interpretable artificial intelligence engine are updated using reinforcement learning methods. Based on the model deviation metrics, the confidence weights of relevant causal rules in the process knowledge graph are adjusted using Bayesian update methods. Based on the overall verification results, the logical rules that map process optimization strategies to control commands in the model factory mechanism are optimized using decision tree induction methods to achieve continuous self-learning and performance evolution of the system. Through the collaborative work of the above six modules, this system establishes a complete technical system from multimodal data synchronous acquisition, dynamic construction and updating of digital twin models, interpretable intelligent deviation tracing, digital space simulation and strategy generation, to online process self-adjustment and closed-loop verification optimization. It realizes accurate perception, intelligent diagnosis, forward-looking optimization and autonomous control of steel structure manufacturing deviations, and provides a complete system implementation for the methods described in claims 1-9.

[0015] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for tracing and self-optimizing manufacturing deviations in steel structures based on multi-source data fusion, characterized in that, Includes the following steps: Multi-modal data from multiple sources are collected during the steel structure manufacturing process, and a digital twin model of the steel structure manufacturing process is constructed and updated synchronously based on the preset process mechanism and the multi-modal data. The multimodal data includes three-dimensional point cloud and visual image data, stress-strain and temperature field data, process timing data and environmental condition data; Based on the digital twin model, the multimodal data is fused, analyzed, and reasoned through an interpretable artificial intelligence engine to generate a deviation source vector with a physical evidence chain. Based on the deviation state represented by the source vector, it is dynamically mapped to the corresponding parameter space of the digital twin model. The deviation source vector includes the root cause process identifier, the associated physical process parameters, and their contribution information. Simultaneously, the deviation source vector is input into the digital twin model that has completed dynamic mapping to perform simulation deduction of deviation propagation and multi-stage root cause decoupling analysis, generating interpretable diagnostic conclusions and corresponding process optimization strategies. Based on the aforementioned process optimization strategy, adaptive control commands are dynamically generated through a model factory mechanism to perform online self-adjustment of the manufacturing process parameters of the manufacturing system. Among them, the manufacturing process status data after the manufacturing process parameters are adjusted is used as new multimodal data input to the digital twin model to verify and iteratively optimize the accuracy of the digital twin model and the process optimization strategy.

2. The method for tracing and self-optimizing manufacturing deviations in steel structures based on multi-source data fusion according to claim 1, characterized in that, The construction and synchronous updating of the digital twin model in the steel structure manufacturing process specifically includes the following steps: Based on the preset process mechanism, an initial digital twin model including a geometric model, a physical model, and a behavioral model is constructed. During the steel structure manufacturing process, the initial digital twin model is synchronously updated using real-time acquired multimodal data, including: based on real-time acquired 3D point cloud and visual image data, the geometric model is registered and compared using a point cloud registration algorithm to generate a geometric deviation field, and the geometric model is corrected in real time in terms of shape and size based on the geometric deviation field. Based on real-time collected stress, strain and temperature field data, the data is used as real-time loads and boundary conditions input to the physical model. The physical model is then subjected to physical simulation through a finite element solver, which drives the real-time updating of the internal mechanical and thermal states of the physical model. Based on real-time collected process timing data and environmental status data, the process logic state and performance evolution path of the behavior model are simulated and updated in real time through rule-based reasoning.

3. The method for tracing and self-optimizing steel structure manufacturing deviations based on multi-source data fusion according to claim 2, characterized in that, The step of registering and comparing the point cloud registration algorithm with the geometric model to generate a geometric deviation field specifically includes the following steps: Extract feature point sets representing the key geometric morphology of steel structure components from 3D point cloud and visual image data, and obtain the corresponding design reference model from the geometric model; The feature point set is matched with the design reference model based on feature descriptors to complete the initial coarse registration. Simultaneously, using the coarse registration result as the initial pose, the 3D point cloud, visual image data, and design reference model are iteratively registered to the nearest point to solve for the optimal spatial transformation matrix. The optimal spatial transformation matrix is ​​then applied to calculate the positional and normal deviations between the 3D point cloud, visual image data, and design reference model, and these deviations are fused to generate a structured geometric deviation field with a continuous spatial distribution.

4. The method for tracing and self-optimizing steel structure manufacturing deviations based on multi-source data fusion according to claim 2, characterized in that, The physical simulation using a finite element method solver, which drives the real-time updating of the internal mechanical and thermal states of the physical model, specifically includes the following steps: The real-time collected stress, strain and temperature field data are correlated with the geometric region corresponding to the physical model to generate the real-time boundary conditions and physical loads required for the finite element simulation model. Based on the real-time boundary conditions and physical loads, the finite element simulation model is solved by thermo-mechanical sequential coupling through the finite element solver, and transient heat conduction analysis and thermo-mechanical analysis are performed in sequence to obtain the transient temperature field, stress field and strain field during the manufacturing process. The transient temperature field, stress field, and strain field are synchronized to the physical model as real-time internal states to drive the updating of the mechanical and thermal fields within the physical model.

5. The method for tracing and self-optimizing steel structure manufacturing deviations based on multi-source data fusion according to claim 2, characterized in that, The real-time inference and update of the process logic state and performance evolution path of the behavior model driven by rule reasoning specifically includes the following steps: The rule parsing engine parses the real-time collected process timing data into specific process instructions and status events, and matches the environmental status data with the preset environmental tolerance threshold in the behavior model to generate environmental trigger events. The process instructions, state events, and environmental trigger events are input into the behavior model. Based on the process rules preset in the behavior model, the input events are matched with the state transition conditions in the preset process rules to determine the current process logic state and deduce the subsequent performance evolution path. Based on the deduction results, the current state nodes and evolution path parameters of the behavior model are updated to achieve real-time logical synchronization with the physical manufacturing process.

6. The method for tracing and self-optimizing manufacturing deviations in steel structures based on multi-source data fusion according to claim 1, characterized in that, The step of using an interpretable artificial intelligence engine to perform fusion analysis and source tracing reasoning on the multimodal data to generate a bias source tracing vector with an attached physical evidence chain specifically includes the following steps: The multimodal data is subjected to spatiotemporal alignment and standardization, and key features are extracted from the multimodal data to construct a unified multidimensional feature vector. The key features include at least: deformation mode statistics extracted from the geometric deviation field, thermo-mechanical coupling feature indices extracted from the stress-strain and temperature field data, and time-series event features extracted from the process time series and environmental state data. The multidimensional feature vectors are input in parallel into the data-driven prediction model and process knowledge graph in the interpretable artificial intelligence engine. The data-driven prediction model is input to obtain preliminary deviation causal hypotheses and their confidence levels. Simultaneously, the process knowledge graph is input to generate candidate causal chains through graph traversal and logical reasoning. The preliminary deviation causal hypotheses and the candidate causal chains are verified and fused to generate a deviation tracing vector containing root cause process identifiers, associated physical process parameters and their contribution levels, and an accompanying physical evidence chain. The verification and fusion at least include: ranking the deviation causal hypotheses according to their confidence levels, and verifying the consistency between the deviation causal hypotheses and the causal chains based on the logical relationships in the process knowledge graph. The physical evidence chain is linked to the multimodal data, deformation mode statistics, thermo-mechanical coupling characteristic indicators, time-series event characteristics, and reasoning paths in the process knowledge graph corresponding to the deviation source vector through reference relationships.

7. The method for tracing and self-optimizing steel structure manufacturing deviations based on multi-source data fusion according to claim 1, characterized in that, The step of inputting the deviation source vector into the dynamically mapped digital twin model to perform deviation propagation simulation and multi-stage root cause decoupling analysis specifically includes the following steps: Based on the deviation tracing vector, the physical process parameters identified by the vector that cause the steel structure manufacturing deviation are located and adjusted in the digital twin model, so as to inject the initial deviation into the corresponding manufacturing process in the digital twin model. Starting from the point where the deviation is injected, the digital twin model performs a multiphysics time-series simulation of the steel structure manufacturing process to deduce the propagation and evolution of the initial deviation along the actual process chain to subsequent stages, and obtains the geometric and mechanical deviations output by each stage during the simulation as the basic data for decoupling analysis. Based on the aforementioned basic data, a contribution decoupling analysis is performed on multiple suspected root causes identified in the deviation tracing vector. Through quantitative calculation, the independent influence and interactive effects of each suspected root cause under the coupling effect of multiple manufacturing stages are distinguished, and a process optimization strategy that clearly points to specific process stages and parameter adjustment directions is generated.

8. The method for tracing and self-optimizing manufacturing deviations in steel structures based on multi-source data fusion according to claim 1, characterized in that, The process of dynamically generating adaptive control commands through a model factory mechanism to adjust the manufacturing process parameters of the manufacturing system online includes the following steps: By receiving and parsing the process optimization strategy, control semantic information about the target process, the parameters to be adjusted, and the adjustment target is extracted. Based on the control semantic information, the pre-set process control logic model in the model factory mechanism is invoked to generate a specific control logic sequence that matches the current manufacturing equipment and operating conditions. The control logic sequence is then mapped to the control interface protocol of the target equipment and encapsulated into a standardized instruction package, which is then transformed into an adaptive control instruction that the manufacturing system can recognize and execute. The adaptive control instruction includes at least real-time parameter settings and action instructions for welding equipment, assembly tooling, and straightening mechanism.

9. The method for tracing and self-optimizing manufacturing deviations in steel structures based on multi-source data fusion according to claim 1, characterized in that, The verification and iterative optimization of the digital twin model and process optimization strategy specifically includes the following steps: The manufacturing process status data generated after executing the adaptive control command will be used as the new multimodal data input to the digital twin model. Forward simulation is performed using the digital twin model, and the simulation prediction results are compared with the corresponding actual results to generate verification results that include a measure of strategy effectiveness and a measure of model bias. Based on the effectiveness measure of the strategy, the weight parameters of the data-driven prediction model in the interpretable artificial intelligence engine are updated using reinforcement learning methods; based on the model bias measure, the confidence weights of relevant causal rules in the process knowledge graph are adjusted using Bayesian update methods; based on the verification results, the logical rules that map process optimization strategies to control commands in the model factory mechanism are optimized using decision tree induction methods.

10. A steel structure manufacturing deviation tracing and self-optimization system based on multi-source data fusion, applied to the method described in any one of claims 1-9, characterized in that, include: The multi-source data acquisition module is used to acquire multimodal data during the steel structure manufacturing process. The multimodal data includes three-dimensional point cloud and visual image data, stress-strain and temperature field data, process timing data and environmental condition data. The digital twin model construction and synchronization module is used to construct and synchronously update a digital twin model of the steel structure manufacturing process based on the preset process mechanism and the multimodal data. An interpretable artificial intelligence engine module is used to perform fusion analysis and source tracing reasoning on the multimodal data, generate a deviation source tracing vector with a physical evidence chain, and dynamically map the deviation state represented by the source tracing vector to the corresponding parameter space of the digital twin model. The simulation and strategy generation module is used to input the deviation source vector into the digital twin model that has completed dynamic mapping, perform simulation and decoupling analysis of deviation propagation and multi-stage root cause analysis, and generate interpretable diagnostic conclusions and corresponding process optimization strategies. The model factory control module is used to dynamically generate adaptive control commands based on the process optimization strategy through the model factory mechanism, and to perform online self-adjustment of the manufacturing process parameters of the manufacturing system. The closed-loop verification and iterative optimization module is used to input the manufacturing process status data after adjusting the manufacturing process parameters as new multimodal data into the digital twin model, and to verify and iteratively optimize the accuracy of the digital twin model and the process optimization strategy.

Images