A digital modeling optimization method for long-span steel anchor box manufacturing process

Abstract

The application discloses a kind of digital modeling optimization methods for large-span steel anchor box manufacturing process.The method generates data source image on GIS vector, CAD drawing, survey database and image data annotation;Collect terminal to obtain paper material, profile, borehole column chart and generate image collection package containing time position;Call data adapter to extract and transcode into unified intermediate representation;The preprocessed data source image and image collection package are input into multi-branch convolutional neural network, and page segmentation, text detection and recognition and key information extraction are performed, and field-level structured results and confidence are output;Based on data source image, field and intermediate representation are aligned and semantic mapping is generated to store standardized data package.Bi-gaussian filter is used for pre-processing, and tau is determined by gray variance mean classification;When the confidence of key field is insufficient or conflicts with image constraints, increase tau and re-filtering to infer and select the optimal output to improve collection efficiency and quality.

Description

Technical Field

[0001] This invention relates to the field of digital modeling technology, and more specifically to a digital modeling optimization method for the manufacturing process of large-span steel anchor boxes. Background Technology

[0002] In long-span bridge engineering, steel anchor boxes, as key load-bearing components of the main cable anchoring system, are typically characterized by large plate thickness, large dimensions, complex structural topology, and numerous and diverse welds. Their manufacturing process often involves multiple collaborative steps, including material preparation and beveling, segmented assembly and positioning, tack welding, multi-layer welding, straightening and machining, non-destructive testing, and final assembly verification. Because the geometric dimensions and welding constraints of steel anchor box components are significantly larger than those of typical box-type components, the assembly-welding coupling effect is prominent. Factors such as fluctuations in heat input, changes in clamping stiffness, and assembly gaps and positioning deviations during manufacturing can cause welding shrinkage and residual stress accumulation, leading to deformation of critical sections, misalignment of the reference chain, and increased risk of weld defects. This results in frequent rework, longer production cycles, increased costs, and difficulty in ensuring consistent quality.

[0003] In existing technologies, the manufacturing process parameters and welding sequence of steel anchor boxes largely rely on the experience of process engineers or are based on a fixed process scheme of a small number of prototypes. This makes it difficult to cover dynamic factors such as differences in material properties between different batches, changes in ambient temperature and humidity, fluctuations in equipment status, and changes in lifting loads. At the same time, clamping strategies such as tooling fixture layout, clamping force, and release timing are often not jointly optimized with the welding sequence and welding parameters, resulting in insufficient deformation control capabilities in high-constraint areas, thick plate areas, or critical reference chain areas. On the other hand, although geometric point cloud data can be obtained on the manufacturing site through laser scanning, photogrammetry, etc., and process parameters such as current, voltage, and welding speed can be collected through welding machine controllers, multi-source data is often stored in a scattered manner, with inconsistent calibers and a lack of unified timestamps and component identification associations. This makes it difficult to form a traceable manufacturing process data package at the process dimension, and even more difficult to establish a queryable process semantic model at the component-weld-station-equipment-fixture level, thus limiting the supporting role of data in process improvement.

[0004] To improve prediction and control capabilities, existing studies have attempted to use finite element simulation to assess welding deformation and residual stress. However, for large-span steel anchor boxes, complete and detailed modeling involves large computational loads, difficult parameter calibration, and long simulation cycles, making it difficult to adapt to actual production cycles. Even with simplified models, prediction deviations are prone to occur due to factors such as assembly constraints and fixture stiffness uncertainties, temporal variations in heat input, and the coupling of thermal effects between adjacent welds. Furthermore, existing data-driven methods often only address a single task (such as deformation prediction or defect identification), lacking multi-task collaborative modeling of key section deformation, key weld quality risks, and residual stress risks. Moreover, the lack of quantification and constraint mechanisms for the uncertainty of prediction results makes it difficult to achieve robust control under the multi-objective trade-off of "quality-cycle time-energy consumption" in optimization decisions. More importantly, existing solutions generally lack a closed-loop mechanism to transform prediction confidence intervals into risk penalty terms and incorporate them into multi-objective optimization. This makes it difficult to achieve collaborative optimization and executable instruction set output for key process elements such as material cutting compensation, segmented assembly sequence, welding sequence, welding parameters, welding vehicle path, and fixture arrangement, under constraints of equipment capability, tooling and fixture accessibility, welding specifications, and cycle time. Furthermore, it is difficult to link instructions, process data, and result data to a knowledge graph to form a traceable data chain to support continuous iteration. Therefore, a digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes is urgently needed to achieve multi-source data alignment and association, semantic modeling of process knowledge graphs, multi-task prediction based on convolutional neural networks, uncertainty-driven risk penalty modeling, and multi-objective optimization and traceability closed loop, thereby improving the stability of steel anchor box manufacturing quality and reducing rework and manufacturing costs. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention discloses a digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes. The method involves collecting geometric point clouds, welding process parameters, environmental parameters, and equipment status parameters corresponding to process nodes; generating process data packages based on unified timestamps and component identifiers; establishing a manufacturing process knowledge graph and mapping the data packages to node and edge attributes to form a process semantic model; extracting geometric semantics, thermal input, assembly constraints, and fixture features into a convolutional neural network prediction model to output key section deformation, weld quality risk, and residual stress risk; performing uncertainty assessment on the output to obtain confidence intervals and generating risk penalty terms; and optimizing at least one of the following under equipment capability and quality constraints: material cutting compensation, welding sequence, welding parameters, welding vehicle path, and fixture arrangement, outputting and executing an executable instruction set; and linking the instruction set and data packages to the knowledge graph to form a traceable data chain.

[0006] A digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes includes the following steps: S1: Collect geometric point cloud data, welding process parameters, environmental parameters, and equipment status parameters corresponding to the process nodes during the manufacturing process of steel anchor boxes, and align and associate them based on a unified timestamp and component identifier to generate a manufacturing process data package; S2: Based on component segmentation, assembly benchmark, weld type and sequence dependency, a knowledge graph of steel anchor box manufacturing process is established, and the manufacturing process data package is mapped to the attribute set of knowledge graph nodes and edges to form a queryable process semantic model. S3: Extract the geometric semantic features, welding thermal input features, assembly constraint features and fixture features of the steel anchor box from the manufacturing process data package, and input them into the trained convolutional neural network prediction model to output the predicted value of the deformation of the key section after welding, the quality risk score of the key weld and the residual stress risk index. S4: Evaluate the uncertainty of the prediction model output to obtain the prediction confidence interval, and generate a risk penalty term based on the prediction confidence interval; S5: Under the conditions of satisfying equipment capacity, tooling and fixture accessibility, welding specifications, cycle time and quality constraints, perform multi-objective optimization on at least one of the following based on risk penalty terms: material cutting compensation, segmented assembly sequence, welding sequence, welding parameters, welding vehicle path and fixture layout, and output and execute an executable instruction set. S6: Link the executable instruction set and manufacturing process data package to the knowledge graph and form a traceable data chain.

[0007] Preferably, the geometric point cloud data includes the point cloud of the plate after blanking and processing, the point cloud of the assembly after segmented assembly, and the point cloud of the steel anchor box after welding, and is registered in the same coordinate system through reference holes and reference surfaces; the welding process parameters include welding current, voltage, welding speed, wire feed speed, number of weld layers, oscillation mode, and welding torch posture; the environmental parameters include ambient humidity and ambient temperature; and the equipment status parameters include equipment power, welding machine duty cycle, welding vehicle running status, lifting load of lifting equipment, assembly gap, positioning deviation, and clamping force.

[0008] Preferably, the alignment association further includes temporal matching of welding process parameters with the geometric point cloud data of the corresponding process node based on the unified timestamp.

[0009] Preferably, the process knowledge graph includes component nodes, weld nodes, assembly station nodes, equipment nodes, and fixture nodes; the edge relationships include assembly dependency, welding sequence dependency, heat-affected coupling, clamping constraints, and station capability constraints; mapping the manufacturing process data package to the attribute set of knowledge graph nodes and edges includes: mapping based on component identifiers, weld identifiers, and process node identifiers, mapping geometric point cloud deviation distribution to component node attributes, mapping welding process parameters to weld node attributes, and mapping assembly constraint parameters and fixture features to attributes of fixture nodes and clamping constraint edges.

[0010] Preferably, the geometric semantic features include at least one of the following: component thickness partitioning features, reinforcing diaphragm topology features, weld adjacency relationship features, and key reference chain features; the key reference chain features are used to characterize the deviation propagation path from the reference hole or reference surface to the key connection surface; the welding heat input features include at least one of the following: heat input per unit length features calculated based on welding electrical parameters and welding speed, heat input fluctuation features, or adjacent weld heat superposition features.

[0011] Preferably, the assembly constraint features include at least one of the following: assembly clearance, positioning deviation, degree-of-freedom constraint matrix, or equivalent parameter of fixture stiffness; the fixture features include at least one of the following: clamping point position, clamping force, release sequence, or fixture accessibility parameter.

[0012] Preferably, the uncertainty assessment includes quantifying at least one of the following: sensor noise, welding heat input fluctuation, clamping force fluctuation, assembly gap fluctuation, or environmental parameter fluctuation, and converting the predicted confidence interval into a penalty term of a multi-objective optimization objective function.

[0013] Preferably, the objectives of the multi-objective optimization include at least two of the following: minimizing the maximum deformation of the critical section, minimizing the rework risk, minimizing the manufacturing cycle, or minimizing energy consumption; the constraints include at least two of the following: equipment capability constraints, accessibility constraints, welding specification constraints, and cycle time constraints; the blanking compensation adopts a zoned compensation strategy: different compensation coefficients are set for the thick plate area, the high-constraint weld neighborhood, or the critical reference chain related area.

[0014] Preferably, the convolutional neural network prediction model is a multi-branch, multi-task convolutional network guided by process relationship priors, comprising: The geometric point cloud encoding branch is used to voxelize or rasterize the geometric point cloud data and input it into a multi-scale 3D convolutional encoder to output multi-resolution geometric semantic features. The welding process timing coding branch is used to assemble welding electrical parameters, welding speed and thermal input features into a timing feature sequence according to the weld bead sequence and input it into a one-dimensional timing convolutional encoder to output a welding thermal input timing representation. The assembly constraint and fixture coding branch is used to construct the assembly clearance, positioning deviation, fixture clamping point position, clamping force and release timing into constraint tensors and input them into the two-dimensional convolutional encoder to output the clamping constraint representation. A relational prior fusion module is used to construct a relation tensor based on weld adjacency relations, thermal influence coupling relations, and key baseline chain relations. It then uses relational convolution and cross-branch attention fusion to jointly encode the geometric semantic features, the welding thermal input temporal representation, and the clamping constraint representation. Conditional normalization or feature modulation mechanisms are employed to adaptively modulate the welding thermal input temporal representation to the geometric semantic features. The multi-task prediction head is used to output the predicted value of post-weld cross-sectional deformation, weld quality risk score, and residual stress risk index respectively.

[0015] This application also provides a digital modeling and optimization system for the manufacturing process of large-span steel anchor boxes, including: The manufacturing process data package generation module collects geometric point cloud data, welding process parameters, environmental parameters, and equipment status parameters corresponding to the process nodes during the manufacturing process of the steel anchor box, and aligns and associates them based on a unified timestamp and component identifier to generate a manufacturing process data package. The steel anchor box manufacturing process knowledge graph establishment module establishes a steel anchor box manufacturing process knowledge graph based on component segmentation, assembly benchmark, weld type and sequence dependency relationship, and maps the manufacturing process data package to the attribute set of knowledge graph nodes and edges to form a queryable process semantic model. The convolutional neural network prediction model calculation module extracts the geometric semantic features, welding thermal input features, assembly constraint features and fixture features of the steel anchor box from the manufacturing process data package, and inputs them into the trained convolutional neural network prediction model, outputting the predicted value of the deformation of the key section after welding, the quality risk score of the key weld and the residual stress risk index. The risk penalty term generation module evaluates the uncertainty of the prediction model output to obtain the prediction confidence interval, and generates a risk penalty term based on the prediction confidence interval. The executable instruction set output and execution module, under the conditions of meeting equipment capacity, tooling and fixture accessibility, welding specifications, cycle time and quality constraints, performs multi-objective optimization on material cutting compensation, segmented assembly sequence, welding sequence, welding parameters, welding car path and fixture layout based on risk penalty items, and outputs and executes the executable instruction set; The recording module links the executable instruction set and manufacturing process data package to the knowledge graph and forms a traceable data chain.

[0016] Compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) This invention collects geometric point clouds, welding process parameters, environmental parameters, and equipment / tooling status parameters at the process node level, and aligns and associates them based on a unified timestamp and component identifier to form a reusable manufacturing process data package, achieving "data homogeneity, consistent standards, and cross-process connectivity". Furthermore, it explicitly models component segmentation, assembly datum, weld type, and sequential dependency relationships as a manufacturing process knowledge graph, and maps the process data package as node and edge attributes to form a queryable process semantic model, transforming key relationships such as assembly dependency, clamping constraints, thermal effect coupling, and station capability from empirical implicitness into searchable and computable explicit information. As a result, high-risk stations and key welds can be located before problems occur, supporting deviation tracing and cause analysis, significantly reducing the delayed correction caused by information fragmentation, improving the transparency of the manufacturing process and the interpretability of quality control, and providing a reliable data foundation for subsequent prediction and optimization.

[0017] (2) This invention explicitly quantifies the uncertainty of the prediction model output into a confidence interval and further transforms it into a risk penalty term, which is introduced into the decision-making process as the objective function term and / or constraint term of multi-objective optimization. This makes the optimization result not only pursue "optimality" but also emphasizes "robust feasibility" under field fluctuations. Under the premise of meeting the constraints of equipment capacity, tooling and fixture accessibility, welding specifications, cycle time and quality, at least one of the following is jointly optimized: material compensation, assembly sequence, welding sequence, welding parameters, welding vehicle path and fixture layout. This can achieve a comprehensive balance of quality, cycle time and energy consumption, and output a set of instructions that can be directly issued and executed, reducing repeated trial welding and rework that rely on manual experience. At the same time, the instruction set, process data package and execution results are associated with the knowledge graph to form a traceable data chain, so that the process scheme can be closed-loop sedimentation and rapid reuse, promoting the transformation of steel anchor box manufacturing from "post-inspection" to "pre-control + process optimization".

[0018] (3) This invention employs a multi-branch, multi-task convolutional neural network guided by process relationship priors: geometric point clouds are extracted with multi-resolution geometric semantic representations through multi-scale three-dimensional convolution; welding electrical parameters and speeds are encoded by one-dimensional temporal convolution according to the weld sequence to obtain thermal input temporal representations; assembly constraints and fixture states are formed by two-dimensional convolution to form clamping constraint representations; and joint encoding is achieved through relation tensors constructed based on weld adjacency, thermal effect coupling, and key benchmark chain relationships under the fusion of relation convolution and cross-branch attention, while conditional normalization / feature modulation is used to adaptively modulate the thermal input temporal to geometric semantics. This structure embeds "topological relationship + process priors" into the network reasoning process, significantly enhancing the ability to express the effects of welding coupling and thermal superposition, and can simultaneously output key cross-section deformation, key weld quality risk, and residual stress risk, reducing the error accumulation caused by multi-model fragmentation; and improving the generalization ability for different configurations, different constraint conditions, and parameter fluctuation scenarios, providing a more reliable prediction basis for subsequent uncertainty assessment and optimization decisions. Attached Figure Description

[0019] Figure 1 This is a flowchart of a digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes according to the present invention; Figure 2 This is a structural diagram of a digital modeling and optimization system for the manufacturing process of large-span steel anchor boxes, as described in this invention. Detailed Implementation

[0020] Those skilled in the art will understand that, in order to make the above-mentioned objects, features, and beneficial effects of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This application presents a digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes, including the following steps: S1: Collect geometric point cloud data, welding process parameters, environmental parameters, and equipment status parameters corresponding to the process nodes during the manufacturing process of steel anchor boxes, and align and associate them based on a unified timestamp and component identifier to generate a manufacturing process data package; In some embodiments, S1 involves the acquisition and alignment of manufacturing process data packages, the definition of process nodes, and the establishment of component identifiers. Taking a large-span steel anchor box as an example, the manufacturing process is divided into several process nodes N1 to N8, such as: N1 blanking and numbering, N2 beveling, N3 segmented assembly and positioning, N4 positioning welding, N5 layered multi-pass welding, N6 post-weld cooling and straightening, N7 machining, and N8 final assembly and inspection. At node N1, a unique component identifier CID (e.g., "project number-segment number-part number-version number") is generated for each plate, partition, stiffener, and segmented assembly based on the design BOM and 3D model. The CID is then fixed in the non-processed area of ​​the component using a high-temperature resistant QR code / steel stamp or RFID. Simultaneously, a binding relationship between the CID and the process node number is established in the MES, serving as the primary key for subsequent data alignment.

[0021] Geometric point cloud data acquisition is implemented by setting up geometric acquisition stations at nodes N1, N3, N6, and N8: N1 (Post-material cutting): A handheld laser scanner or photogrammetry system is used to quickly scan the outer contour, key holes, and reference edges of the sheet metal. After point cloud acquisition, a "sheet metal point cloud file + key feature point set" is automatically generated. N3 (Post-assembly assembly): A fixed laser tracker or a mobile 3D scanning vehicle is deployed around the assembly jig to scan the reference surfaces, mating edges, and key connection surfaces of the segmented assembly; simultaneously, the coordinates of the jig's reference points are acquired to achieve registration in the same coordinate system. N6 (Post-weld and post-correction): Key sections (such as anchorage areas, web-top and bottom plate junctions, and areas with dense stiffening ribs) are re-measured, and post-weld morphology point clouds and a "key section deviation report" are output. N8 (Final assembly inspection): Global point cloud acquisition is performed on the key reference chain of the final assembly (reference holes / reference surfaces to key connection surfaces) to form the final inspection point cloud. During point cloud acquisition, the acquisition station reads the component CID (code scanning / RFID) and records the acquisition process node, scanning equipment number and operator number in the acquisition software to generate point cloud metadata.

[0022] Welding process parameter acquisition is implemented, with automatic data collection at nodes N4 and N5: welding current, voltage, wire feed speed, welding mode, duty cycle, etc., are read from the welding machine controller / welding power supply; travel speed, welding torch posture, trajectory segment number, and weld bead number are read from the welding vehicle or robot control system; the start and end times, start and end positions (which can be provided by the welding vehicle odometer / positioning system), and process parameter curves of each weld bead are packaged and saved with "weld number WID - weld bead number BID" as the granularity. To ensure that the weld bead corresponds to the component, a "weld list" is issued by the process system before welding, and each weld bead is bound to the component CID and process node; at the start of welding, the welder selects WID / BID on the terminal or automatically loads it by scanning a code, and the process parameter data of the weld bead is automatically transmitted back after welding is completed.

[0023] Environmental and equipment status parameter collection is implemented by deploying environmental and equipment status collection units at each key workstation: Environmental parameters: Temperature and humidity sensors are deployed at assembly / welding workstations, sampling at fixed intervals to form time series; if necessary, indicators such as wind speed / dust can be added to evaluate the effectiveness of protection. Equipment status parameters: Welding machine duty cycle, alarm codes, and cooling status are collected; welding vehicle operating status (start / stop, speed, mileage), power or supply status, and fault codes are collected; lifting equipment at lifting nodes (such as N3 segmented lifting positioning) collects lifting load, hook height / amplitude, and key action time points to record assembly posture and assembly disturbances. The above parameters are also bound to process node numbers and workstation numbers, and timestamps are recorded uniformly.

[0024] To ensure the fusion of multi-source data, a unified alignment rule is established for the alignment of timestamps and component identifiers: 1) All acquisition devices are connected to the factory time server (e.g., NTP / PTP) to record data with a unified timestamp; offline devices undergo time calibration during data import. 2) The alignment primary key adopts "CID + process node number + time window," where the time window is determined by the start / end event of the process. For example: point cloud acquisition event: the scan start / end time is used as the window; weld event: the weld start and end time is used as the window; environment and equipment status: the time series segment overlapping with the above window is taken. 3) "WID / BID" is added as a secondary primary key for weld / weld bead level data to ensure traceability to specific weld locations and process execution details. 4) The point cloud coordinate system is unified: the reference hole / reference surface or jig reference point is used as the registration reference to transform point clouds at different stages to the same factory coordinate system or component coordinate system for subsequent extraction of "geometric semantic features" and "critical section deviations."

[0025] After data collection and alignment, a manufacturing process data package P(CID,Ni) is generated for each component CID at each process node. This package includes at least the following: Identification information: project number, CID, process node number, workstation number, timestamp range, operator / equipment number; Geometric data: point cloud file path or point cloud summary, key feature points, key section deviation entries, registration reference information; Welding data: WID / BID list, parameter curve summary for each weld pass, abnormal events (such as arc interruption, parameter out-of-bounds); Environmental data: temperature / humidity time series summary and extreme values ​​within the window; Equipment status: welding machine / welding vehicle / lifting equipment status sequence summary, alarm code and occurrence time. After generation, the data package is written to the database and sent back to the MES. Simultaneously, the data package index (CID, process node, WID / BID) is used for subsequent steps S2 knowledge graph mapping and S3 feature extraction.

[0026] S2: Based on component segmentation, assembly benchmark, weld type and sequence dependency, a knowledge graph of steel anchor box manufacturing process is established, and the manufacturing process data package is mapped to the attribute set of knowledge graph nodes and edges to form a queryable process semantic model. In some embodiments, the modeling objects and input data sources for the atlas are exemplified by a steel anchor box for a long-span bridge. The steel anchor box consists of a top plate, bottom plate, web plate, diaphragm, stiffening ribs, etc., and is divided into several segments according to the manufacturing strategy (e.g., left anchor segment, right anchor segment, intermediate transition segment, etc.). Each segment is further subdivided into plates and sub-assemblies. The inputs for atlas construction include: 1) Design-side data: 3D model, BOM, segment / component list, datum hole / datum surface definition, weld list (weld number WID, type, location, length, associated components), process route and process node definition; 2) Manufacturing-side data: manufacturing process data package P(CID,Ni) generated by S1, which includes point cloud deviation, welding process curves, environment and equipment status, etc.

[0027] 2. Definition of Node Types in the Knowledge Graph (Minimum Usable Set): Establish a knowledge graph G=(V,E) for the steel anchor box manufacturing process, where the node set V contains at least the following five categories: Component Nodes Vc: Representing plates / partitions / stiffening ribs / segment assemblies, with node attributes including CID, material, plate thickness partitioning, key reference feature set, geometric point cloud summary, deviation statistics, etc.; Weld Nodes Vw: Representing specific weld WID and its weld bead BID, with node attributes including weld type (butt weld / fillet weld, etc.), location, length, design requirements, weld bead sequence, welding parameter curve summary, abnormal events, and quality. Risk scoring, etc.; Process node Vp: represents process node Ni (material cutting, assembly, tack welding, multi-pass welding, straightening, final inspection, etc.), node attributes include start / end time, workstation number, cycle time requirement, actual time taken, environmental statistics, etc.; Equipment node Ve: represents welding machine, welding vehicle / robot, lifting equipment, scanning equipment, etc., node attributes include equipment number, status sequence, alarm code, duty cycle, availability index, etc.; Fixture node Vf: represents jig / fixture / clamping device, node attributes include clamping point location, reachability parameters, clamping force curve, release sequence, stiffness equivalent parameters, etc. Note: If clamping force / release sequence data collection is not available on-site, "fixture configuration ID + clamping point geometry" can be used as the initial attribute, and iterated and improved later without affecting the establishment of the map framework.

[0028] Edge relationship definition: How segments, datums, types, and order dependencies are plotted. The edge set E must include at least the following relationships (each edge can have attributes): 1) Segment / containment relationship, PART_OF(CID_child → CID_parent): The panel belongs to the sub-assembly, and the sub-assembly belongs to the segment assembly; Edge attributes: assembly tolerance requirements, assembly datum reference, etc. 2) Assembly datum relationship (making the "assembly datum" explicit), REFER_BASE(CID → BaseID): The component references a datum hole / datum surface (BaseID can be used as a datum node or as a component node attribute); BASE_CHAIN(BaseID_a → BaseID_b): Datum chain transmission relationship, used to express the deviation propagation path from "datum hole / datum surface to critical connection surface"; Edge attributes: datum type, measurement method, allowable deviation threshold. 3) Weld connection and type relationships: JOIN_BY_WELD(CID_i ↔ CID_j, WID): Two components are connected by a certain weld; WELD_TYPE(WID → Type): Weld type relationship (butt weld, fillet weld, plug weld, etc.); Edge attributes: weld location, design level, bevel type, heat input limit, etc. 4) Sequence dependency relationships: ASSEMBLY_PRECEDENCE(CID_a → CID_b): Assembly sequence dependency; WELD_PRECEDENCE(WID_m → WID_n): Weld sequence dependency; PROCESS_ROUTE(CID / WID → Ni): Component / weld is associated with a process node (in which process it is completed); Edge attributes: dependency reason (location requirements / constraint requirements / accessibility), suggested window, allowed insertion range, etc. 5) Thermal-affected zone coupling and clamping constraint relationship, THERMAL_COUPLING(WID_m ↔ WID_n): Thermal-affected zone coupling between adjacent welds; edge attributes include coupling distance and thermal superposition intensity level; CLAMP_CONSTRAINT(Vf → CID): Clamping constraint of the fixture on the component; edge attributes include clamping point set, clamping force range, release sequence, etc. 6) Workstation capacity / equipment capacity constraint relationship, WORKSTATION_CAPABILITY(Ni → Ve): Matching of process node with equipment capacity; edge attributes include accessibility, maximum plate thickness capacity, tooling interface, available time period, etc.

[0029] The mapping rules for "manufacturing process data package → node / edge attribute set" are as follows for the data package P(CID,Ni) generated by S1: 1) Locate the mapping primary key: use "CID+Ni" as the primary key to determine the component node Vc and the process node Vp; if the data package contains the weld / weld bead identifier "WID / BID", then further locate the weld node Vw; locate the equipment node Ve with the equipment number; locate the fixture node Vf with the fixture configuration ID. 2) Point cloud deviation mapping: write the deviation statistics after point cloud registration (e.g., key reference surface flatness, hole offset, key section contour deviation) into the attributes of the component node Vc; at the same time, write the "reference chain related deviation" into the BASE_CHAIN ​​or REFER_BASE edge attribute for subsequent deviation propagation query. 3) Welding curve and abnormal event mapping: write the curve summary of welding current / voltage / speed, the number of parameter out-of-bounds times, the number of arc interruptions, etc. into the attribute of the weld node Vw; write the process node where welding occurred into the PROCESS_ROUTE(WID → Ni) edge attribute, and record the start and end time windows. 4) Environment and equipment status mapping: Temperature and humidity statistics within the Ni window are written to the Vp attribute of the process node (or the weld node as context); equipment status sequences and alarm codes are written to the Ve attribute of the equipment node, and the equipment availability index for the current process is recorded on the WORKSTATION_CAPABILITY side. 5) Clamping constraint mapping (if clamping force / release sequence is collected): Clamping force curves and release sequences are written to the Vf attribute of the fixture node, and the clamping strategy for the current operation is recorded on the CLAMP_CONSTRAINT(Vf → CID) side. After mapping, a "process semantic model" is formed: It can not only query data by component, weld, and process, but also trace semantic relationships such as "install before welding," "thermal coupling neighborhood," and "reference chain propagation" along dependency edges.

[0030] Typical queries for the process semantic model are available. Query 1: Locate the neighborhood of a high-risk deformation weld. Input: A component's CID shows excessive deviation at a critical section after welding. Query: Find the associated weld JOIN_BY_WELD from the component node Vc, then extend it along THERMAL_COUPLING to the set of adjacent welds. Return the heat input fluctuations, equipment alarms, and environmental temperature and humidity statistics for these welds at the corresponding process nodes to explain the source of the deviation. Query 2: Trace the propagation path of the baseline chain deviation. Input: A critical connection surface offset was found during final inspection. Query: Find the propagation path from the baseline hole / baseline to the connection surface along BASE_CHAIN. Combine this with the point cloud deviation attributes at each process node to determine whether the deviation was introduced during assembly positioning or post-weld shrinkage. Query 3: Process reuse. Input: New batch isomorphic segmentation. Query: Match the historical graph subgraph according to the component topology and weld type, directly reuse the parameter window, fixture strategy, and sequence dependency of the corresponding weld to form a reusable process template.

[0031] Incremental updates and version management of the data map: Whenever a new data packet P(CID,Ni) is generated, the system performs an incremental update: if the node / edge does not exist, it is created; if it exists, the attribute version is appended according to the timestamp; key attributes (point cloud deviation, welding curve summary, alarm code) are archived in the form of "batch version number + time window"; a "traceable data chain" is formed: the executable instruction set ID and the data packet ID are jointly attached to the process node Vp and the weld node Vw to ensure that the process instructions, execution process and equipment status can be traced back from the results.

[0032] S3: Extract the geometric semantic features, welding thermal input features, assembly constraint features and fixture features of the steel anchor box from the manufacturing process data package, and input them into the trained convolutional neural network prediction model to output the predicted value of the deformation of the key section after welding, the quality risk score of the key weld and the residual stress risk index. In some embodiments, S3 feature extraction and convolutional neural network prediction output, input data determination and sample granularity, taking the steel anchor box segment assembly CID=SA-03 as the object, trigger a prediction once after the process node N5 (layered multi-pass welding). The system reads the following content from the manufacturing process data package: Point cloud data: N3 post-assembly point cloud, N5 post-weld point cloud (or only post-weld point cloud if available), and point cloud registration reference (reference hole / reference surface); Welding process data: the set of weld numbers (WID) associated with this segment and its weld sequence (BID), including current, voltage, welding speed, wire feed speed, oscillation mode, welding torch posture, arc interruption and boundary crossing events for each weld; Environment and equipment / tooling status: temperature and humidity sequence within the welding window, welding machine duty cycle / alarm code, welding vehicle running status; If clamping force / release timing, assembly gap / positioning deviation, etc. are collected, they are used as assembly constraints and fixture status inputs; Relationship priors (from the S2 process semantic model): adjacency relationships, heat-affected coupling relationships, welding sequence dependency relationships, critical reference chain relationships, and fixture clamping constraint relationships on components between SA-03 associated welds. The sample granularity for prediction adopts a "two-level parallel" approach: 1) Segment-level samples: used to output the predicted values ​​of deformation of key sections and residual stress risk indicators; 2) Weld-level samples: used to output the quality risk scores of key welds (a score is given for each WID, and then aggregated into the segment-level risk overview).

[0033] Geometric semantic feature extraction (from point cloud) (1) Point cloud cleaning and registration: outlier removal, unified scale and normal estimation of N3 and N5 point clouds; rigid registration based on reference holes / reference surfaces to unify the point cloud to the component coordinate system. (2) Semantic partitioning and structural topology coding: the segment is divided into several geometric semantic regions: thick plate region, stiffening rib dense area, partition plate neighborhood, weld neighborhood, key reference chain neighborhood (located by reference chain relationship). Generate geometric description for each region: local curvature / flatness trend, hole position offset trend, key section contour deviation distribution, etc., and record the adjacency relationship between the region and the weld WID. (3) Voxelization / rasterization to form geometric input tensor: voxelize the registered point cloud at a fixed resolution (or project the key section into a multi-channel raster) to form a three-dimensional geometric tensor; at the same time, write "region label, plate thickness partition, reference chain neighborhood mark" as additional channels to obtain "geometric input with semantic channels". Output: Multi-resolution geometric semantic features input (for use in 3D convolution branches).

[0034] Welding heat input feature extraction (from weld bead timing) (1) Weld bead sequence construction: According to the welding sequence dependency in the process semantic model, sort the weld beads (BID) corresponding to the weld seam of SA-03 to form one or more timing segments (e.g., divided by weld seam group, by work station, by continuous operation window). (2) Heat input and stability feature encoding: Extract the following for each weld bead: electrical parameter statistics (mean, fluctuation, number of times of exceeding the limit, number of times of arc interruption); speed and wire feeding stability statistics; welding torch posture change amplitude; heat input characterization (heat input trend per unit length formed by the combination of electrical parameters and speed), and heat superposition prompt mark between adjacent weld beads (given by the heat influence coupling relationship). (3) Forming one-dimensional timing input: Concatenate the above weld bead features in time sequence into a one-dimensional multi-channel sequence as "heat input timing input". Output: Welding heat input timing feature input (for use by 1D timing convolution branch).

[0035] Assembly constraint features and fixture feature extraction (from tooling / assembly status), (1) Assembly constraint features: read assembly gap, positioning deviation, and assembly datum consistency index from the manufacturing process data package; if on-site inspection is used instead (such as back-calculation of gap / deviation from point cloud after assembly), then the N3 point cloud features are back-inferred and written into the feature set. (2) Fixture features: read the fixture clamping point position (relative to the component coordinate system), clamping force curve summary, release sequence, and fixture reachability parameters; if only the fixture configuration ID is available, then map it to the pre-calibrated fixture parameter template. (3) Two-dimensional constraint tensor construction: project "clamping point distribution, clamping force level, release sequence, gap / deviation partition value" onto the two-dimensional mesh of the key area of ​​the component to form a multi-channel two-dimensional tensor that reflects the spatial distribution of clamping constraints. Output: Assembly constraint and fixture feature input (for use in 2D convolution branch).

[0036] The relational prior tensor (derived from the process semantic model) generates relational prior inputs based on edge relations from the knowledge graph, guiding cross-branch fusion: Adjacency relations: weld adjacency matrix (which welds are spatially adjacent / share components); Thermal influence coupling relations: thermal coupling strength level (weak / medium / strong) and range of action; Baseline chain relations: region mask covered by critical baseline chains (which geometric regions are sensitive to critical connection surfaces); Clamping constraint relations: constraint connections between clamps and component regions (clamping point-region mapping). The above relational information is organized into a "relationship tensor / relationship mask" readable by the network and aligned with geometric / temporal / constraint features.

[0037] The convolutional neural network prediction model employs a reasoning process of "multi-branch encoding + relational prior fusion + multi-task output": 1) Geometric point cloud encoding branch (3D CNN): A three-dimensional geometric tensor is input into a multi-scale three-dimensional convolutional encoder to obtain multi-layer geometric semantic representations from coarse to fine, used to capture overall deformation trends and local warping features. 2) Welding process temporal encoding branch (1D TCN): The thermal input temporal sequence is input into a one-dimensional temporal convolutional encoder to obtain temporal representations reflecting "thermal input fluctuations, superposition effects, and abnormal events." 3) Assembly constraint and fixture encoding branch (2D CNN): A two-dimensional constraint tensor is input into a two-dimensional convolutional encoder to extract the spatial distribution representation of clamping constraints, highlighting the influence of high-constraint regions and release temporal changes on deformation. 4) Relationship Prior Fusion Module: Input the relationship tensor into the fusion module and perform two types of fusion: Relationship Convolution: Aggregate weld-level features based on weld adjacency and thermal coupling relationships to obtain a weld representation that "considers the influence of adjacent welds"; Cross-Branch Attention Fusion: Guided by the baseline chain region mask, adaptively modulate the geometric representation with the temporal thermal input representation and clamping constraint representation (e.g., through conditional normalization / feature modulation, to make thermal input changes produce stronger feature responses in key areas). 5) Multi-Task Prediction Head Output: Key Section Deformation Prediction Value: Output deformation index for a specified set of key sections (defined by the baseline chain and key connection surfaces) to form a section-level prediction table; Key Weld Quality Risk Score: Output a risk score of 0–1 or 0–100 for the set of key welds and provide a Top-K list of high-risk welds; Residual Stress Risk Index: Output a segmented risk index (which can be used for subsequent S4 uncertainty assessment and S5 optimization).

[0038] After inference is completed, the system generates a prediction result object R(CID,N5), including: R_deform: key section ID → predicted deformation value (and corresponding section location); R_weld_risk: WID → quality risk score (and marks the risk source characteristics, such as thermal input fluctuation / strong clamping restraint / strong thermal coupling); R_stress_risk: segmented residual stress risk index; and writes R into the extended field of the manufacturing process data package, while writing the key weld risk score back into the corresponding weld node attribute in the knowledge graph, providing input for uncertainty assessment in S4 and multi-objective optimization in S5.

[0039] Training samples and annotation sources: The convolutional neural network prediction model is trained offline, and the training data comes from historical batches of manufacturing process data packages and inspection results: the deformation annotation of key sections comes from the deviation report of post-weld point cloud and final inspection point cloud; the weld quality risk annotation comes from non-destructive testing results (such as ultrasonic / radiographic / magnetic particle) and rework records; the residual stress risk annotation can come from strain / displacement monitoring, process test data of key areas, or calibrated simulation / empirical benchmarks as weak annotations.

[0040] S4: Evaluate the uncertainty of the prediction model output to obtain the prediction confidence interval, and generate a risk penalty term based on the prediction confidence interval; In some embodiments, S4 uncertainty assessment, confidence interval generation, and risk penalty term construction, input objects, and assessment triggering timing, taking segmented assembly CID=SA-03 as an example, after process node N5 (layered multi-pass welding), the system obtains the prediction result R(CID,N5) output by S3, which includes: a set of predicted values ​​for critical section deformation: providing predicted deformation values ​​for each critical section SID; a set of risk scores for critical weld quality: providing risk scores for each critical weld WID; and a segmented residual stress risk index: providing segmented risk indices. Simultaneously, sensor noise indicators, thermal input fluctuation statistics, clamping force fluctuation statistics, assembly gap fluctuations, environmental temperature and humidity fluctuations, and equipment alarms from the manufacturing process data packet P(CID,N5) are read as input evidence for uncertainty assessment.

[0041] Uncertainty sources are modeled, and the system decomposes uncertainties into two categories: 1) Data uncertainty: caused by sensor noise, sampling gaps, time synchronization errors, point cloud registration errors, and welding electrical parameter sampling jitter; 2) Process uncertainty: caused by fluctuations in welding heat input, clamping force and release timing, assembly gap changes, environmental temperature and humidity fluctuations, and equipment status drift (excessive duty cycle, alarm code triggering), etc. To facilitate engineering implementation, the system sets quantifiable indicators for each type of uncertainty, such as: point cloud registration residual level, welding current / voltage fluctuation level, clamping force curve stability level, temperature and humidity fluctuation amplitude level, equipment alarm level, etc., and uniformly normalizes them into an "uncertainty evidence vector" of 0 to 1.

[0042] The uncertainty assessment method involves performing multiple inferences on a convolutional neural network prediction model to estimate the distribution characteristics of the model output. The specific process is as follows: Without changing the main content of the input data packet, a random perturbation mechanism is enabled for the model inference process, such as enabling random deactivation during the inference phase in several convolutional blocks or fusion modules of the network; the inference is repeated several times to obtain multiple sets of deformation prediction values ​​for each key cross-section SID, multiple sets of risk scores for each key weld WID, and multiple sets of residual stress risk indicators; based on the above multiple sets of results, the output dispersion is calculated to characterize the uncertainty of the model itself; simultaneously, an "uncertainty evidence vector" is introduced to weight and correct the output dispersion: when a high level of thermal input fluctuation, a high level of clamping force fluctuation, or a high level of point cloud registration residual is detected, the system increases the uncertainty weight of that batch of outputs to reflect the impact of on-site fluctuations on prediction reliability. In some embodiments, an "input perturbation" method can also be used instead of "inference perturbation," for example, multiple samples of thermal input features and clamping force features within their allowable fluctuation range are taken, and then multiple inferences are performed to obtain the distribution; both can form confidence intervals.

[0043] The system generates confidence intervals for predictions (providing intervals for three types of outputs). After obtaining the output distribution characteristics, the system generates confidence intervals for different outputs: Confidence interval for critical section deformation: For each critical section (SID), outputs a "predicted deformation interval [lower bound, upper bound]", and records the interval width as the predicted stability index for that section; Confidence interval for critical weld risk score: For each weld (WID), outputs a "risk score interval [lower bound, upper bound]". When the upper bound of the interval is high and the interval width is large, the weld is determined to be "high-risk and uncertain"; Confidence interval for residual stress risk index: Outputs segmented risk intervals and uses them as "conservative constraints" in subsequent optimization. Furthermore, the system performs rule-based processing on the confidence intervals: If there are missing input data or equipment alarms are triggered, the corresponding output interval is expanded to the preset worst-case level to ensure a safe decision boundary.

[0044] The risk penalty term is constructed by transforming the confidence interval into a risk penalty term Penalty(CID, N5) for S5 multi-objective optimization. This penalty term consists of three parts: For each critical section SID, the allowable deviation threshold of that section is read (from design / inspection specifications or process semantic model). If the upper bound of its confidence interval is close to or exceeds the threshold, a penalty value is generated. The "amount by which the upper bound exceeds the threshold" and the "interval width" are comprehensively weighted, so that: the closer the prediction is to the deviation, the greater the penalty; the higher the uncertainty (the wider the interval), the greater the penalty (reflecting the robustness requirement).

[0045] For high-risk weld quality, a weld-level penalty value is generated if the upper limit of the risk score range of a critical weld (WID) exceeds a preset risk threshold. For welds belonging to a strong heat-affected zone or located in a high-constraint region, the weight is increased to reflect the amplification effect of prior process relationships on quality risk.

[0046] For residual stress risk penalties, if the upper bound of the confidence interval of a segmented residual stress risk index exceeds a threshold, a segmented penalty value is generated. This penalty is then used as a constraint-type penalty (i.e., a risk item that leans more towards "hard constraints") to suppress welding sequences and clamping strategies that may lead to high residual stress accumulation. Finally, the system synthesizes the above three types of penalties with configurable weights to form a unified risk penalty item Penalty(CID,N5), while retaining the individual penalty items to facilitate S5's trade-offs among different objectives.

[0047] The system writes the uncertainty assessment results in a structured manner: for each SID: deformation prediction value, confidence interval, interval width, and corresponding penalty value; for each WID: risk score, confidence interval, interval width, and corresponding penalty value; for segments: residual stress risk index, confidence interval, and corresponding penalty value; and uses Penalty(CID,N5) as the input parameter set for S5 multi-objective optimization. At the same time, the confidence interval and penalty value of the key weld are written back to the weld node attributes of the knowledge graph, forming a traceable "prediction-uncertainty-decision" link, providing a data foundation for parameter self-calibration of subsequent batch optimization templates.

[0048] S5: Under the conditions of satisfying equipment capacity, tooling and fixture accessibility, welding specifications, cycle time and quality constraints, perform multi-objective optimization on at least one of the following based on risk penalty terms: material cutting compensation, segmented assembly sequence, welding sequence, welding parameters, welding vehicle path and fixture layout, and output and execute an executable instruction set. In some embodiments, based on the multi-objective optimization and executable instruction set generation execution of risk penalty items, the input preparation is optimized. (1) Basic process scheme: The process system reads the predetermined process route and initial scheme of the segment, including: weld list WID and weld BID, default welding sequence (by weld number), default welding parameter window (by weld type / plate thickness partition to give current, voltage, speed and other ranges), default fixture configuration (fixture ID and clamping point set), and default walking path of welding vehicle (by station layout to give coarse path). (2) Risk penalty items and key risk objects: The system obtains Penalty(CID,N5) from S4, which includes at least: "deformation penalty" and "confidence interval width" (uncertainty) of key section SID; "quality risk penalty" of key weld WID (especially welds with high upper limit of interval); segment residual stress risk penalty (with greater weight for high restraint areas). At the same time, it reads from the process semantic model of S2: weld adjacency, heat-affected coupling relationship, welding sequence dependence, key reference chain area mask, fixture accessibility and station capability constraints, etc. (3) Equipment and on-site resource constraints: The system reads on-site resources: available welding machine / welding vehicle (or robot) number, current available time period, capacity boundaries of each equipment (maximum plate thickness capacity, available welding mode), upper limit of station cycle time, number of fixtures and available fixture configuration, welding gun posture / accessibility restrictions, welding specification requirements (e.g., upper and lower limits of parameters for each type of weld, preheating / interpass temperature rules, allowed layer combinations, etc.).

[0049] In this embodiment, the optimization variables are defined as follows: 1) Welding sequence variables: Weld seam level sequence: sorting and grouping of WIDs (e.g., rearranging according to symmetrical welding, skip welding, segmented back welding strategies); Weld bead level sequence: layer combination sequence and alternation strategy of the same weld seam (e.g., short weld bead first then long weld bead, fillet weld first then butt weld, or vice versa, adjusted according to coupling relationship). 2) Welding parameter variables: setting specific values ​​for welding current, voltage, welding speed, wire feed speed, oscillation mode, etc. for different weld seam types / plate thickness zones (within the allowable range of specifications); setting more conservative parameter windows or smoother heat input change strategies for weld seams in strong thermal coupling neighborhoods. 3) Welding vehicle path variables: path sequence, transfer strategy and attitude switching strategy of the welding vehicle accessing weld seam groups within the workstation (ensuring accessibility and avoiding interference); path linkage strategy for prioritizing or delaying welding of high-risk weld seams (linked with sequence variables). 4) Fixture layout variables: clamping point activation / deactivation combination, clamping force level (or target range), release sequence (release non-critical areas first, release reference chain areas later, etc.); strategies to apply higher constraints or release later to critical reference chain related areas to reduce reference drift risk. Optional extension: If at nodes N1 to N3, this step can also incorporate "material cutting compensation (zoning compensation coefficient / rules)" and "segmented assembly sequence (sub-assembly assembly sequence and positioning strategy)" into the same optimization framework. This embodiment highlights the S5 closed loop and selects key variables from the welding stage for demonstration.

[0050] Multi-objective and constraint modeling (incorporating risk penalty terms into optimization), (1) Optimize objective settings. The system simultaneously minimizes the following indicators in a multi-objective manner: Objective A: Deformation risk of critical sections (summed by S4 deformation penalty terms, with higher weight for key sections); Objective B: Quality risk of critical welds (summed by S4 weld penalty terms, with higher weight for thermally coupled strong neighborhoods); Objective C: Residual stress risk (given by S4 residual stress penalty terms, serving as a "hard constraint" objective); Objective D: Manufacturing cycle (welding time + transfer time + fixture change time, meeting the cycle time limit); Objective E: Total energy consumption or heat input (used to suppress deformation accumulation caused by excessive heat input). (2) Constraints: Equipment capacity constraints: The parameters of each weld must fall within the range achievable by the equipment and allowed by the welding specifications; Tooling and fixture accessibility constraints: The clamping point of the fixture should not interfere with the path of the welding torch / welding cart, and the posture of the welding torch should meet the accessibility requirements; Welding specification constraints: The layer combination corresponding to the weld type, heat input restrictions, and necessary waiting / cooling windows; Cycle time constraints: The total process time should not exceed the upper limit of the workstation cycle time; Quality constraints: Solutions with the upper limit of the risk scoring interval of the critical weld exceeding the threshold are directly judged as infeasible or subject to high penalties; Solutions with the upper limit of the confidence interval of the deformation of the critical section exceeding the threshold are directly judged as infeasible or subject to high penalties. Solution process (layered optimization, taking into account feasibility and efficiency) is adapted to the production cycle time.

[0051] This embodiment adopts a hierarchical optimization process of "discrete priority, continuous refinement": Step 4.1: Candidate Solution Generation (Discrete Layer). The system first generates several feasible candidate solutions based on the process semantic model: multiple welding sequences that satisfy the dependencies are generated according to the welding sequence dependency relationship; for weld pairs with strong heat-affected coupling, different strategy versions such as "alternating welding / symmetrical welding / skipped welding / retreating welding" are generated; multiple reachable paths for welding vehicles are generated according to the workstation layout (e.g., "nearest first, then farthest", "routing by zone", "reducing the number of transfers"); and several fixture arrangement and release timing schemes are generated according to the fixture template library (e.g., "release after the critical reference chain area" and "segmented release in high-constraint areas"). Each candidate solution forms a "discrete strategy combination".

[0052] Step 4.2: Rapid Evaluation and Screening (Intra-Model Evaluation). For each discrete strategy combination, the system automatically constructs its corresponding input features and calls the S3 prediction model to obtain three types of prediction outputs. Then, S4 generates confidence intervals and risk penalty terms to form the target vector (deformation / mass / stress / cycle / energy consumption) of the scheme. The system eliminates schemes that violate hard constraints and retains several Pareto-optimal candidate sets for the next step.

[0053] Step 4.3: Parameter Refinement (Continuous Layers). For the selected candidate set, the system performs local search and refinement of welding parameters (current, voltage, speed, oscillation mode, and layer combination values) without changing the discrete order / path / fixture topology. For high-risk welds, priority is given to reducing parameter fluctuations, lowering peak heat input, increasing inter-layer waiting time, or changing layer combinations. For welds in critical cross-section sensitive areas, priority is given to parameter combinations that result in a more significant reduction in deformation penalty terms. While ensuring cycle time, energy consumption and total heat input are considered. After each parameter update, S3+S4 is called again to calculate the penalty terms until the preset number of iterations is reached or the target improvement tends to stabilize.

[0054] Step 4.4: Optimal Solution Selection. From the Pareto candidate set, select the final execution plan according to the project strategy: for example, prioritizing quality and deformation, followed by cycle time, or increasing the cycle time weight if the delivery time for a particular batch is tight. The final output is a set of determined parameters: welding sequence, welding parameters, welding vehicle path, fixture arrangement / release timing (and optional material cutting compensation and assembly sequence).

[0055] The generated executable instruction set, ultimately compiled into an executable instruction set InstructionSet_ID, includes at least: 1) Clamping instructions: fixture configuration ID, clamping point list, target force range / level for each clamping point, clamping sequence; release sequence: which clamping points are released after which weld groups are completed, with clamping points in critical reference chain areas being released last; interference check results and safety warnings. 2) Welding operation instructions: weld / bead operation queue: arranged according to the optimized WID / BID order, and labeled with symmetrical welding / skip welding / retreat welding strategy parameters; parameter table for each weld bead: current, voltage, speed, wire feed, oscillation mode, welding torch posture requirements, allowable fluctuation range; process monitoring thresholds for critical welds: such as threshold for the number of times electrical parameters exceed limits, arc interruption threshold, inter-layer waiting requirements, etc. 3) Welding vehicle path instructions: access path sequence within the workstation, transfer nodes, posture switching points; entry / exit posture constraints for corresponding weld groups to avoid interference between fixtures and components. 4) Online verification and triggering conditions (reserved for S6 traceability and subsequent closed loop): trigger a point cloud retest or key benchmark check after a specific weld group is completed; if the risk score of a key weld increases during execution (e.g., abnormal parameter fluctuations are detected in real-time monitoring), trigger a pause and switch to the backup plan or re-optimize.

[0056] The instruction set is executed and recorded. The instruction set is issued to the welding machine / welding vehicle controller and workstation terminal through MES: the welding vehicle moves automatically according to the path instructions and executes the welding queue; the welding machine outputs closed-loop control according to the parameter table; the tooling fixtures execute according to the clamping and releasing sequence (which can be driven by the controller or guided by the workstation terminal for manual execution and confirmation); during the execution, welding curves, equipment status and environmental data are collected in real time to generate new incremental manufacturing process data packages; the InstructionSet_ID, execution log, process data package and detection results are written back to the knowledge graph to form a traceable data chain, providing reusable process templates and optimized initial values ​​for the next batch.

[0057] S6: Link the executable instruction set and manufacturing process data package to the knowledge graph and form a traceable data chain.

[0058] In some embodiments, the S6 instruction set is associated with the manufacturing process data package to form a traceable data chain. The traceability object and unique identifier are designed. Taking the segmented assembly CID=SA-03 as an example, after the process node N5 (layered multi-pass welding) is completed, the system obtains: the executable instruction set InstructionSet_ID=IS-2026-03-03-N5-001 (generated by S5 and issued for execution); the manufacturing process data package P(CID,Ni): includes the N3 post-assembly point cloud package, the N5 welding process package, the N6 post-weld retest point cloud package, etc.; execution log and results: welding curve, equipment alarm, environmental sequence, key section retest deviation report, non-destructive testing results, and rework record. To ensure consistent traceability across systems, a three-level primary key system is adopted: 1) Component primary key: CID (unique identifier for component / segment); 2) Process primary key: Ni (process node number) + Workstation_ID (workstation number); 3) Job primary key: WID / BID (weld / bend number) or Scan_ID (scan task number). Simultaneously, a Run_ID (e.g., RUN-2026-03-03-0007) is generated for each "optimization-execution" closed loop as a unified session identifier for instruction sets and data packet increments. In the knowledge graph's connection structure (nodes / edges and attributes), based on the process knowledge graph established in S2, two new types of nodes are added and key edge relationships are supplemented to meet traceability requirements: (1) New node types: Instruction Set Node Vi: Represents an executable instruction set; attributes include InstructionSet_ID, generation time, optimization target weight, constraint set summary, version number, distribution object (equipment / workstation), applicable component CID, applicable process Ni, etc. Data Packet Node Vd: Represents a manufacturing process data packet or its incremental packet; attributes include DataPack_ID, acquisition time window, data type (point cloud / welding curve / environment / equipment status / inspection report), storage address summary, data integrity verification value, etc.

[0059] (2) Add / supplement edge relationships: APPLY_TO(Vi → CID): The instruction set is applied to a component / segment; EXECUTE_AT(Vi → Ni): The instruction set is executed at a process node; GENERATE(Vi → Vd): The incremental data packet generated by the execution of the instruction set (execution log / curve / retest, etc.); EVIDENCE_OF(Vd → WID / BID): The data packet corresponds to a specific weld / weld bead (if applicable); MEASURE_OF(Vd → BaseID / SID): The point cloud / inspection data corresponds to the baseline chain or critical section (SID is the critical section ID); CAUSE_LINK(Vd ↔ Ve / Vf): The data packet is associated with the equipment node / fixture node (e.g., alarm code, clamping force curve). Through the above nodes and edges, a four-element closed-loop structure of "instruction-execution-data-object" is formed in the graph.

[0060] Association Rules: How to "automatically attach" instruction sets and data packets to the graph? The system performs association according to the following rules: Rule 1: Instruction set attachment. When S5 generates InstructionSet_ID, the system reads the applicable objects in it: CID, Ni, WID / BID queue, fixture configuration ID, welding vehicle path ID, parameter table version, etc., and creates a new instruction set node Vi in the graph; establishes APPLY_TO (Vi → CID) and EXECUTE_AT (Vi → Ni) edges; writes "weld / bead queue summary, fixture clamping / releasing timing summary, welding parameter window summary, welding vehicle path summary" into the Vi attribute or writes the edge attribute of Vi to the corresponding weld node Vw (for traceability accurate to the weld level). Rule 2: Data packet attachment. During execution, S1 collects and forms or incrementally updates the manufacturing process data packet DataPack_ID. The system reads CID, Ni, time window, equipment number, WID / BID (if any), and Scan_ID from the data packet metadata, and creates or updates the data packet node Vd in the graph; establishes the GENERATE (Vi → Vd) edge (automatically associated by matching Run_ID with the time window); if the data contains WID / BID, then establishes the EVIDENCE_OF (Vd → WID / BID) edge or associates it with the weld node Vw; if the data packet is a point cloud / section deviation report, then establishes the MEASURE_OF (Vd → SID / BaseID) edge and writes the deviation information into the edge attribute.

[0061] Rule 3: Device / Fixture Association. Establish CAUSE_LINK or corresponding attribute association between the device alarm code, duty cycle, welding vehicle status, clamping force curve, etc. in the data packet and the device node Ve and fixture node Vf for subsequent "cause localization tracing".

[0062] The structured output of the traceable data chain, after being connected, forms the following data chain for CID=SA-03 in process N5: CID=SA-03 → (EXECUTE_AT) → Ni=N5 → (association) → Vi=IS-2026-03-03-N5-001 → (GENERATE) → Vd_1 (welding curve package) → Vd_2 (equipment status package) → Vd_3 (environment package) → Vd_4 (post-weld point cloud package) → Vd_5 (non-destructive testing report). At the same time, each data packet node is further linked to the weld WID / BID and critical section SID through EVIDENCE_OF or MEASURE_OF, realizing fine-grained traceability from "segment level" to "weld / section level".

[0063] Typical traceability query implementation method, query example 1: From the final inspection deviation, trace back the process instructions and execution evidence. When the final inspection finds that the deformation of the critical section SID-07 exceeds the tolerance: Locate the data packet node Vd (post-weld point cloud / final inspection report) corresponding to MEASURE_OF (Vd → SID-07) in the map; trace back along the GENERATE edge to the instruction set node Vi to obtain the current welding sequence, welding parameter table, and fixture release sequence; then locate the weld WID set adjacent to SID-07 along EVIDENCE_OF, read its welding curve packet Vd_1 and alarm packet Vd_2, and output "which weld experienced heat input fluctuation / arc interruption / equipment alarm at what time", to achieve interpretable traceability. Example 2: Locating why a high-risk weld is classified as high-risk. For a weld with a high WID-23 risk score: trace its corresponding weld curve package Vd, and retrieve the number of parameter out-of-bounds occurrences and the level of heat input fluctuation; simultaneously retrieve the curve packages of its heat-affected coupling neighboring welds to determine if there is heat superposition; output the associated fixture constraint edge attributes (clamping force / release timing) as influencing factors to form a "risk explanation list". Example 3: Process reuse and comparison. For new batch isomorphic segments, query the subgraphs in the historical graph that are topologically similar to them, extract the instruction set node Vi and its key parameter range that have the "lowest deformation penalty and satisfy the cycle time" as the initial solution for S5 optimization, improving optimization efficiency.

[0064] To ensure reliable traceability and version management and data integrity, this embodiment adopts the following mechanisms: Version number mechanism: Instruction set Vi records the version number and the reason for the change (e.g., "parameter downgrade triggered by device alarm"); Data packet Vd records the data version and the firmware version of the acquisition device; Integrity verification: A verification digest is calculated for each data packet and written to the Vd attribute; A verification digest is calculated for the instruction set content (parameter table, path, fixture strategy) and written to the Vi attribute; Chain association: Using Run_ID as the session identifier, all Vds generated in the same optimization-execution are chained together to avoid cross-batch confusion and enhance the consistency and auditability of the data chain.

[0065] In some embodiments, the geometric point cloud data includes the point cloud of the plate after blanking and processing, the point cloud of the assembly after segmented assembly, and the point cloud of the steel anchor box after welding, and is registered in the same coordinate system through reference holes and reference surfaces; the welding process parameters include welding current, voltage, welding speed, wire feed speed, number of weld layers, oscillation mode, and welding torch posture; the environmental parameters include ambient humidity and ambient temperature; and the equipment status parameters include equipment power, welding machine duty cycle, welding vehicle running status, lifting load of lifting equipment, assembly gap, positioning deviation, and clamping force.

[0066] In some embodiments, the alignment association further includes temporal matching of welding process parameters with geometric point cloud data of the corresponding process node based on the unified timestamp.

[0067] In some embodiments, the process knowledge graph includes component nodes, weld nodes, assembly station nodes, equipment nodes, and fixture nodes; edge relationships include assembly dependencies, welding sequence dependencies, heat-affected coupling, clamping constraints, and station capability constraints; mapping the manufacturing process data package to the attribute set of knowledge graph nodes and edges includes: mapping based on component identifiers, weld identifiers, and process node identifiers, mapping geometric point cloud deviation distribution to component node attributes, mapping welding process parameters to weld node attributes, and mapping assembly constraint parameters and fixture features to attributes of fixture nodes and clamping constraint edges.

[0068] In some embodiments, the geometric semantic features include at least one of the following: component thickness partitioning features, reinforcing diaphragm topology features, weld adjacency relationship features, and key reference chain features; the key reference chain features are used to characterize the deviation propagation path from the reference hole or reference surface to the key connection surface; the welding heat input features include at least one of the following: heat input per unit length features calculated based on welding electrical parameters and welding speed, heat input fluctuation features, or adjacent weld heat superposition features.

[0069] In some embodiments, the assembly constraint features include at least one of the following: assembly clearance, positioning deviation, degree-of-freedom constraint matrix, or equivalent parameter of fixture stiffness; the fixture features include at least one of the following: clamping point location, clamping force, release sequence, or fixture accessibility parameter.

[0070] In some embodiments, uncertainty assessment includes quantifying at least one of sensing noise, welding heat input fluctuation, clamping force fluctuation, assembly gap fluctuation, or environmental parameter fluctuation, and converting the predicted confidence interval into a penalty term of a multi-objective optimization objective function.

[0071] In some embodiments, the objectives of the multi-objective optimization include at least two of the following: minimizing the maximum deformation of the critical section, minimizing the rework risk, minimizing the manufacturing cycle, or minimizing energy consumption; the constraints include at least two of the following: equipment capability constraints, accessibility constraints, welding specification constraints, and cycle time constraints; the blanking compensation adopts a zoned compensation strategy: different compensation coefficients are set for the thick plate area, the high-constraint weld neighborhood, or the critical reference chain related area.

[0072] In some embodiments, the convolutional neural network prediction model is a multi-branch, multi-task convolutional network guided by process relationship priors, including: The geometric point cloud encoding branch is used to voxelize or rasterize the geometric point cloud data and input it into a multi-scale 3D convolutional encoder to output multi-resolution geometric semantic features. The welding process timing coding branch is used to assemble welding electrical parameters, welding speed and thermal input features into a timing feature sequence according to the weld bead sequence and input it into a one-dimensional timing convolutional encoder to output a welding thermal input timing representation. The assembly constraint and fixture coding branch is used to construct the assembly clearance, positioning deviation, fixture clamping point position, clamping force and release timing into constraint tensors and input them into the two-dimensional convolutional encoder to output the clamping constraint representation. A relational prior fusion module is used to construct a relation tensor based on weld adjacency relations, thermal influence coupling relations, and key baseline chain relations. It then uses relational convolution and cross-branch attention fusion to jointly encode the geometric semantic features, the welding thermal input temporal representation, and the clamping constraint representation. Conditional normalization or feature modulation mechanisms are employed to adaptively modulate the welding thermal input temporal representation to the geometric semantic features. The multi-task prediction head is used to output the predicted value of post-weld cross-sectional deformation, weld quality risk score, and residual stress risk index respectively.

[0073] In some embodiments, a multi-branch, multi-task convolutional network guided by process relationship priors is used. Taking the steel anchor box segment assembly CID=SA-03 as an example, an inference is triggered after process node N5 (layered multi-pass welding). The system reads the following from the manufacturing process data package: Geometric point cloud: post-assembly point cloud (N3) and post-weld point cloud (N5 or N6), and reference hole / reference surface registration information; Welding process: the weld bead sequence BID corresponding to each weld WID and its current, voltage, welding speed, wire feed speed, oscillation mode, welding torch posture, and abnormal events; Assembly and fixtures: assembly gap, positioning deviation, fixture clamping point position, clamping force curve summary, and release sequence; Process relationship priors: weld adjacency relationships, thermal effect coupling relationships, and critical reference chain coverage areas derived from the knowledge graph. The output of the network inference includes: post-weld critical section deformation prediction value, critical weld quality risk score, and residual stress risk index (recommended to be consistent with the claims, all limited by "critical"). Geometric point cloud encoding branch (multi-scale 3D convolutional encoder), (1) Point cloud preprocessing: outlier removal and registration of the point cloud, unified to the component coordinate system. (2) Voxelization / rasterization: the point cloud is divided into a 3D voxel grid, and each voxel records the occupancy, local height statistics, normal / curvature summary, etc.; at the same time, "thickness partition label, weld neighborhood label, and reference chain sensitive area label" are written as additional channels into the voxel. (3) Multi-scale 3D convolutional encoding: a hierarchical convolutional structure from coarse to fine is used to extract features: the coarse-scale layer is used to capture the segmented overall warping trend and the overall contraction trend; the fine-scale layer is used to capture the local deformation of the weld neighborhood and the local offset of the reference chain neighborhood; the output of each layer forms a multi-resolution geometric semantic feature set, providing "global + local" geometric representation for subsequent fusion.

[0074] Welding process timing coding branch (one-dimensional timing convolutional encoder), (1) Weld sequence construction: Based on the welding sequence dependency of the process knowledge graph, the welds are arranged according to the actual execution order or the planned order; for welds across workstations, they can be segmented according to the continuous operation window. (2) Timing feature extraction: For each weld, a feature vector is generated, which includes at least electrical parameter statistics, speed statistics, thermal input characterization, number of parameter out-of-bounds times, number of arc interruptions, and inter-layer waiting time. (3) 1D timing convolutional coding: The weld feature sequence is input into the one-dimensional timing convolutional encoder to obtain "thermal input timing characterization", which is used to characterize the trend of thermal input fluctuation, thermal superposition, and abnormal events on deformation and quality. (4) Thermal coupling prompt: For weld pairs with strong thermal influence coupling, a "coupling mark" channel is added to the sequence to enable the encoder to perceive the coupling relationship and superposition risk between adjacent welds.

[0075] Assembly constraint and fixture coding branch (2D convolutional encoder), (1) Constraint tensor construction: Taking the component unfolding plane or key section projection as a reference, the assembly gap and positioning deviation are mapped to a spatial distribution grid; the fixture clamping point position is mapped to a sparse heat map channel; the clamping force level and release timing are mapped to the corresponding channel's numerical or sequential mark. (2) 2D convolutional coding: The above multi-channel 2D constraint tensor is input into the 2D convolutional encoder to extract the clamping constraint characterization, so that the network can identify the spatial coupling between the high-constraint region, the release sensitive region and the reference chain sensitive region.

[0076] The relation prior fusion module (relation tensor + relation convolution + cross-branch attention) (1) Relation tensor construction: Weld adjacency relation tensor: indicates whether welds are adjacent, share component boundaries or spatial distance levels; thermal influence coupling relation tensor: indicates the thermal coupling strength level and range of action between weld pairs; key reference chain relation tensor: marks "regions / weld sets sensitive to the propagation of deviations of key connection surfaces" in the form of a mask. (2) Relation convolution aggregation: aggregate weld-level features (weld neighborhood features from temporal and geometric branches) according to adjacency and thermal coupling relationships to obtain an enhanced representation that "considers the influence of adjacent welds", which is used to suppress misjudgments caused by isolated modeling of single welds. (3) Cross-branch attention fusion: using the reference chain relation tensor as attention guide, align and fuse geometric semantic features, thermal input temporal representation and clamping constraint representation, and prioritize enhance the feature response of the reference chain sensitive area and the thermally coupled strong neighborhood. (4) Conditional normalization / feature modulation: The thermal input time sequence is used as the modulation signal to adaptively modulate the geometric branch features: When the thermal input fluctuation increases or is superimposed and enhanced, the network's response to the geometric features in the weld neighborhood and the sensitive area of ​​the reference chain is amplified, thus better conforming to the physical law of welding thermal deformation.

[0077] The multi-task prediction head (shared representation + separate outputs) inputs the fused shared representation into the multi-task prediction head, which outputs: 1) Post-weld critical section deformation prediction values: outputs section-level deformation indices for a predefined set of critical sections (determined by the baseline chain and structural features), and can simultaneously output section location indices; 2) Critical weld quality risk score: outputs risk scores and risk levels (e.g., low / medium / high) for critical weld WIDs, used to guide subsequent optimization and process monitoring; 3) Residual stress risk index: outputs segmented risk indices or risk levels for critical areas, used for uncertainty assessment and robust optimization. During the training phase, the three tasks share encoder parameters and are optimized separately on their respective prediction heads to achieve collaborative learning of "deformation-quality-stress," reducing error accumulation caused by the fragmentation of multiple models.

[0078] This network integrates multi-source information to improve the comprehensiveness and consistency of predictions. It jointly encodes three types of heterogeneous information—point cloud geometry, welding thermal input timing, and clamping constraints—within the same framework. Through a multi-task prediction head, it simultaneously outputs key section deformation, key weld quality risk, and residual stress risk. This avoids the inconsistencies in feature caliber and error propagation caused by the separation of "deformation model, quality model, and stress model" in existing technologies. It significantly enhances the overall characterization of the steel anchor box welding coupling problem, providing consistent and comparable prediction outputs for subsequent uncertainty assessment and multi-objective optimization. By introducing prior knowledge of process relationships, the network enhances its ability to express and generalize coupling effects. It constructs a relationship tensor through weld adjacency, thermal effect coupling, and key baseline chain relationships, and uses relationship convolution and cross-branch attention to achieve prior-guided feature fusion. The network can explicitly model key mechanisms such as "thermal superposition of adjacent welds, sequential dependence, and baseline chain deviation propagation", thereby reducing the risk of accidental correlation and overfitting caused by relying solely on data-driven approaches. It can still maintain good generalization performance under scenarios of configuration changes, constraint condition changes, or parameter fluctuations, and improve the accuracy of identifying highly constrained regions and key baseline chain sensitive regions. Heat input-driven feature modulation makes inference more consistent with engineering principles and enhances the ability to identify key risks. By employing conditional normalization or feature modulation mechanisms, the temporal representation of heat input is adaptively modulated to geometric features. This allows the network to automatically strengthen the response of the weld neighborhood and sensitive areas of the baseline chain when heat input fluctuations increase and coupling intensifies, thereby more sensitively capturing potential deformation deviations and quality defect risks. Combined with clamping constraint branches to model the spatial distribution of clamping force and release timing, it can more accurately distinguish between situations such as "heat-induced deformation dominance" and "springback dominance caused by constraint release," improving the reliability of key weld risk scores and residual stress risk indicators, and providing a more robust decision-making basis for process optimization.

[0079] This application also provides a digital modeling and optimization system for the manufacturing process of large-span steel anchor boxes, such as... Figure 2As shown, the system hardware consists of the following components: Firstly, the sensing and acquisition hardware includes: 1) Geometric acquisition equipment (point cloud), fixed 3D laser scanner / laser tracker (deployed at assembly stations, post-weld retesting stations, and final inspection stations), mobile point cloud acquisition vehicle or handheld laser scanner (used for retesting of sheet metal and local areas after material cutting), and station reference calibration components (reference hole / reference surface calibration targets, positioning ball targets, etc.); 2) Welding process acquisition equipment, welding machine power supply / welding machine controller (providing current, voltage, mode, duty cycle, etc.), welding vehicle / welding robot controller (providing speed, trajectory segment, welding torch posture, BID / WID, alarm, etc.), and optional: weld seam tracking sensors (laser vision tracking, arc voltage tracking) to supplement welding torch offset and tracking quality; 3) Tooling fixture and assembly status acquisition equipment, and fixture clamping force sensor. 4) Equipment and environmental status acquisition devices, including: hydraulic / pneumatic pressure sensors or tension / compression sensors; clamp position / stroke sensors (displacement switches, encoders) for recording clamping point status and release sequence; assembly gap / positioning deviation measurement devices (digital feeler gauges, displacement gauges, laser displacement sensors; or calculated from post-assembly point clouds); 5) Identification and traceability acquisition devices, including: RFID readers / QR code scanners (for binding CID, WID / BID, process node Ni); and label printing equipment (high-temperature resistant QR code label printers / steel stamping equipment). (Note: The text also mentions various equipment and environmental status acquisition devices, such as temperature and humidity sensors (placed near assembly and welding stations), wind speed / dust sensors (for assessing protective gas / site disturbances), and equipment status acquisition (welding machine alarms, welding vehicle operating status, power / supply status, lifting equipment load sensors, etc.). Edge computing and data aggregation hardware (workstation side): 1) Industrial data acquisition gateway, equipped with multiple protocol interfaces: Industrial Ethernet, RS485, CAN, DI / DO, analog signals, supporting industrial protocols such as OPC UA / Modbus TCP / RTU, Profinet / EtherNet / IP, used for unified sampling, buffering, timestamp alignment and preliminary cleaning of welding machine / welding vehicle / fixture / environmental sensor data; 2) Edge inference server, industrial computer / rack server (including GPU or NPU accelerator card), used for point cloud preprocessing (denoising, registration, voxelization), feature extraction and convolutional neural network inference (S3), and can complete multiple inferences for preliminary uncertainty estimation locally; 3) Time synchronization device, factory clock source / time server (NTP or PTP Grandmaster), switch supporting PTP (optional), used to ensure the "unified timestamp" requirement. Central computing and storage hardware (workshop / plant data center): 1) Application server cluster, process knowledge graph server (graph database / knowledge graph engine), optimization computing server (CPU-based, optional GPU for accelerated evaluation; running multi-objective optimization and instruction compilation), traceability and auditing server (logs, versioning, signature / verification); 2) Storage system, object storage / file server (stores large files such as point cloud files, original welding curve records, and inspection reports), relational database server (stores CID, process, instruction set index, and metadata), time-series database server (stores high-frequency time-series data such as temperature and humidity, welding parameter curves, and equipment status), backup storage / disaster recovery equipment (NAS / tape library / off-site backup optional); 3) Network and security equipment, core switches, industrial switches, routers, firewalls, industrial isolation gateways / security gateways (optional, used for isolation between production and office networks). Execution control hardware (distribution and closed-loop): 1) MES / workstation terminal (HMI), industrial touch screen / industrial tablet (for displaying instruction sets, confirming clamping steps, and handling abnormalities), MES interface server (for work orders, process node Ni, and resource status management); 2) Controllers and actuators, welding machine controller, welding vehicle / robot controller (receiving welding parameters and path instructions), fixture control unit (PLC + solenoid valve / servo valve / hydraulic station), receiving clamping force settings and release timing, lifting equipment control interface (optional, for hoisting event alignment and safety interlocking). Hardware connection methods (typical topology and data flow), workshop network topology, production control network (OT network): industrial switches form a star or ring network (redundant RSTP / ERPS), data service network (IT network): core switches converge to the data center, OT and IT are isolated by firewalls / gateways, only necessary ports are opened and one-way / two-way policies are applied. Connections from field devices to the edge gateway: 1) Scanner / tracker → Edge inference server, connection method: Gigabit / 10 Gigabit Ethernet (RJ45 / fiber optic), direct transmission of large file point clouds, synchronization method: read CID and Ni and write metadata when the scan task is triggered; timestamp from NTP / PTP; 2) Welding machine / welding vehicle controller → Industrial data acquisition gateway, connection method: Industrial Ethernet (OPC UA / Modbus TCP / Profinet, etc.), data content: electrical parameters, speed, attitude, alarm, weld start and end events, WID / BID; 3) Fixture sensor / valve island / PLC → Industrial data acquisition gateway, connection method: RS485 (Modbus RTU), DI / DO, analog, or industrial Ethernet, data content: clamping force, clamping point status, release sequence, stroke / position; 4) Environmental sensor → Industrial data acquisition gateway, connection method: RS485 / LoRa / Wi-Fi (depending on site wiring conditions), ultimately aggregated to the gateway, data content: temperature and humidity time series, fluctuation statistics, 5) RFID / barcode scanner → workstation HMI / edge gateway, connection method: USB / serial port / Bluetooth (connect to the nearest workstation terminal), the terminal writes CID, WID / BID, Ni The connection from the edge to the center side involves an industrial data acquisition gateway to a central time-series database / message bus (optional), connected via Ethernet; reporting sampled data and event data (weld start / end, alarms, etc.); an edge inference server to a central application server / object storage, connected via Ethernet; uploading point cloud processing results, features, model output, and confidence intervals; and a time synchronization server to all network devices, connected via NTP / PTP broadcast or unicast; key devices (gateway, welding machine, welding vehicle, inference server) must be synchronized. The instruction set issuance and execution are connected in a closed loop: Central optimization server (S5) → MES interface server → workstation HMI / controller. The issued content includes: welding sequence queue, welding parameter table, welding vehicle path, and fixture clamping / releasing sequence. Workstation HMI / PLC → Fixture actuator (valve island / hydraulic station) executes: clamping force setting and release action. Status feedback is used for closed-loop recording. Welding vehicle / robot controller → Welding machine controller coordinates: trajectory and parameter linkage execution; process curves and alarms are fed back. The topology diagram illustrates the connection relationships as follows: 1) The point cloud acquisition device is connected to the edge inference server via Ethernet; the edge inference server is connected to the object storage / application server via a core switch. 2) The welding machine controller, welding vehicle controller, fixture PLC, and environmental sensors are connected to the industrial data acquisition gateway via industrial Ethernet or RS485; the industrial data acquisition gateway is connected to the time-series database server and knowledge graph server via Ethernet. 3) The time synchronization server is connected to the industrial data acquisition gateway, edge inference server, welding machine controller, welding vehicle controller, and workstation HMI via NTP / PTP to achieve unified timestamps. 4) The optimization server issues executable instruction sets to the welding vehicle / robot controller, welding machine controller, and fixture PLC via the MES interface; execution process data is transmitted back from each controller to the industrial data acquisition gateway and written to the storage system; the knowledge graph server associates the instruction set ID, data packet ID, and CID / Ni / WID / BID to form a traceable data chain.

[0080] The manufacturing process data package generation module collects geometric point cloud data, welding process parameters, environmental parameters, and equipment status parameters corresponding to the process nodes during the manufacturing process of the steel anchor box, and aligns and associates them based on a unified timestamp and component identifier to generate a manufacturing process data package. The steel anchor box manufacturing process knowledge graph establishment module establishes a steel anchor box manufacturing process knowledge graph based on component segmentation, assembly benchmark, weld type and sequence dependency relationship, and maps the manufacturing process data package to the attribute set of knowledge graph nodes and edges to form a queryable process semantic model. The convolutional neural network prediction model calculation module extracts the geometric semantic features, welding thermal input features, assembly constraint features and fixture features of the steel anchor box from the manufacturing process data package, and inputs them into the trained convolutional neural network prediction model, outputting the predicted value of the deformation of the key section after welding, the quality risk score of the key weld and the residual stress risk index. The risk penalty term generation module evaluates the uncertainty of the prediction model output to obtain the prediction confidence interval, and generates a risk penalty term based on the prediction confidence interval. The executable instruction set output and execution module, under the conditions of meeting equipment capacity, tooling and fixture accessibility, welding specifications, cycle time and quality constraints, performs multi-objective optimization on material cutting compensation, segmented assembly sequence, welding sequence, welding parameters, welding car path and fixture layout based on risk penalty items, and outputs and executes the executable instruction set; The recording module links the executable instruction set and manufacturing process data package to the knowledge graph and forms a traceable data chain.

[0081] Compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) This invention collects geometric point clouds, welding process parameters, environmental parameters, and equipment / tooling status parameters at the process node level, and aligns and associates them based on a unified timestamp and component identifier to form a reusable manufacturing process data package, achieving "data homogeneity, consistent standards, and cross-process connectivity". Furthermore, it explicitly models component segmentation, assembly datum, weld type, and sequential dependency relationships as a manufacturing process knowledge graph, and maps the process data package as node and edge attributes to form a queryable process semantic model, transforming key relationships such as assembly dependency, clamping constraints, thermal effect coupling, and station capability from empirical implicitness into searchable and computable explicit information. As a result, high-risk stations and key welds can be located before problems occur, supporting deviation tracing and cause analysis, significantly reducing the delayed correction caused by information fragmentation, improving the transparency of the manufacturing process and the interpretability of quality control, and providing a reliable data foundation for subsequent prediction and optimization.

[0082] (2) This invention explicitly quantifies the uncertainty of the prediction model output into a confidence interval and further transforms it into a risk penalty term, which is introduced into the decision-making process as the objective function term and / or constraint term of multi-objective optimization. This makes the optimization result not only pursue "optimality" but also emphasizes "robust feasibility" under field fluctuations. Under the premise of meeting the constraints of equipment capacity, tooling and fixture accessibility, welding specifications, cycle time and quality, at least one of the following is jointly optimized: material compensation, assembly sequence, welding sequence, welding parameters, welding vehicle path and fixture layout. This can achieve a comprehensive balance of quality, cycle time and energy consumption, and output a set of instructions that can be directly issued and executed, reducing repeated trial welding and rework that rely on manual experience. At the same time, the instruction set, process data package and execution results are associated with the knowledge graph to form a traceable data chain, so that the process scheme can be closed-loop sedimentation and rapid reuse, promoting the transformation of steel anchor box manufacturing from "post-inspection" to "pre-control + process optimization".

[0083] (3) This invention employs a multi-branch, multi-task convolutional neural network guided by process relationship priors: geometric point clouds are extracted with multi-resolution geometric semantic representations through multi-scale three-dimensional convolution; welding electrical parameters and speeds are encoded by one-dimensional temporal convolution according to the weld sequence to obtain thermal input temporal representations; assembly constraints and fixture states are formed by two-dimensional convolution to form clamping constraint representations; and joint encoding is achieved through relation tensors constructed based on weld adjacency, thermal effect coupling, and key benchmark chain relationships under the fusion of relation convolution and cross-branch attention, while conditional normalization / feature modulation is used to adaptively modulate the thermal input temporal to geometric semantics. This structure embeds "topological relationship + process priors" into the network reasoning process, significantly enhancing the ability to express the effects of welding coupling and thermal superposition, and can simultaneously output key cross-section deformation, key weld quality risk, and residual stress risk, reducing the error accumulation caused by multi-model fragmentation; and improving the generalization ability for different configurations, different constraint conditions, and parameter fluctuation scenarios, providing a more reliable prediction basis for subsequent uncertainty assessment and optimization decisions.

[0084] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products, and therefore this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0085] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes, characterized in that, Including the following steps: S1: Collect geometric point cloud data, welding process parameters, environmental parameters, and equipment status parameters corresponding to the process nodes during the manufacturing process of steel anchor boxes, and align and associate them based on a unified timestamp and component identifier to generate a manufacturing process data package; S2: Based on component segmentation, assembly benchmark, weld type and sequence dependency, a knowledge graph of steel anchor box manufacturing process is established, and the manufacturing process data package is mapped to the attribute set of knowledge graph nodes and edges to form a queryable process semantic model. S3: Extract the geometric semantic features, welding thermal input features, assembly constraint features and fixture features of the steel anchor box from the manufacturing process data package, and input them into the trained convolutional neural network prediction model to output the predicted value of the deformation of the key section after welding, the quality risk score of the key weld and the residual stress risk index. S4: Evaluate the uncertainty of the prediction model output to obtain the prediction confidence interval, and generate a risk penalty term based on the prediction confidence interval; S5: Under the conditions of satisfying equipment capacity, tooling and fixture accessibility, welding specifications, cycle time and quality constraints, perform multi-objective optimization on at least one of the following based on risk penalty terms: material cutting compensation, segmented assembly sequence, welding sequence, welding parameters, welding vehicle path and fixture layout, and output and execute an executable instruction set. S6: Link the executable instruction set and manufacturing process data package to the knowledge graph and form a traceable data chain.

2. The digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes according to claim 1, characterized in that, The geometric point cloud data includes the point cloud of the plate after blanking and processing, the point cloud of the assembly after segmented assembly, and the point cloud of the steel anchor box after welding, and is registered in the same coordinate system through reference holes and reference surfaces; the welding process parameters include welding current, voltage, welding speed, wire feed speed, number of weld layers, oscillation mode, and welding torch posture; the environmental parameters include ambient humidity and ambient temperature; and the equipment status parameters include equipment power, welding machine duty cycle, welding vehicle running status, lifting load of lifting equipment, assembly gap, positioning deviation, and clamping force.

3. The digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes according to claim 1, characterized in that, Alignment association also includes temporal matching of welding process parameters with the geometric point cloud data of the corresponding process nodes based on the unified timestamp.

4. The digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes according to claim 1, characterized in that, The process knowledge graph includes component nodes, weld nodes, assembly station nodes, equipment nodes, and fixture nodes; edge relationships include assembly dependencies, welding sequence dependencies, heat-affected coupling, clamping constraints, and station capability constraints. Mapping the manufacturing process data package to the attribute set of knowledge graph nodes and edges includes: mapping based on component identifiers, weld identifiers, and process node identifiers; mapping geometric point cloud deviation distribution to component node attributes; mapping welding process parameters to weld node attributes; and mapping assembly constraint parameters and fixture features to the attributes of fixture nodes and clamping constraint edges.

5. The digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes according to claim 1, characterized in that, The geometric semantic features include at least one of the following: component thickness partitioning features, reinforcing diaphragm topology features, weld adjacency relationship features, and key reference chain features; the key reference chain features are used to characterize the deviation propagation path from the reference hole or reference surface to the key connection surface; the welding heat input features include at least one of the following: heat input per unit length features calculated based on welding electrical parameters and welding speed, heat input fluctuation features, or adjacent weld heat superposition features.

6. The digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes according to claim 1, characterized in that, The assembly constraint features include at least one of the following: assembly clearance, positioning deviation, degree-of-freedom constraint matrix, or equivalent parameter of fixture stiffness; the fixture features include at least one of the following: clamping point position, clamping force, release sequence, or fixture accessibility parameter.

7. The digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes according to claim 1, characterized in that, Uncertainty assessment includes quantifying at least one of the following: sensor noise, welding heat input fluctuation, clamping force fluctuation, assembly gap fluctuation, or environmental parameter fluctuation, and converting the predicted confidence interval into a penalty term of a multi-objective optimization objective function.

8. The digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes according to claim 1, characterized in that, The objectives of the multi-objective optimization include at least two of the following: minimizing the maximum deformation of the critical section, minimizing the rework risk, and minimizing the manufacturing cycle or energy consumption; the constraints include at least two of the following: equipment capacity constraints, accessibility constraints, welding specification constraints, and cycle time constraints. The blanking compensation adopts a zoned compensation strategy: different compensation coefficients are set for thick plate areas, high-constraint weld neighborhoods, or critical reference chain related areas.

9. The digital modeling and optimization method for the manufacturing process of large-span steel anchor boxes according to claim 1, characterized in that, The convolutional neural network prediction model is a multi-branch, multi-task convolutional network guided by process relationship priors, including: The geometric point cloud encoding branch is used to voxelize or rasterize the geometric point cloud data and input it into a multi-scale 3D convolutional encoder to output multi-resolution geometric semantic features. The welding process timing coding branch is used to assemble welding electrical parameters, welding speed and thermal input features into a timing feature sequence according to the weld bead sequence and input it into a one-dimensional timing convolutional encoder to output a welding thermal input timing representation. The assembly constraint and fixture coding branch is used to construct the assembly clearance, positioning deviation, fixture clamping point position, clamping force and release timing into constraint tensors and input them into the two-dimensional convolutional encoder to output the clamping constraint representation. A relational prior fusion module is used to construct a relation tensor based on weld adjacency relations, thermal influence coupling relations, and key baseline chain relations. It then uses relational convolution and cross-branch attention fusion to jointly encode the geometric semantic features, the welding thermal input temporal representation, and the clamping constraint representation. Conditional normalization or feature modulation mechanisms are employed to adaptively modulate the welding thermal input temporal representation to the geometric semantic features. The multi-task prediction head is used to output the predicted value of post-weld cross-sectional deformation, weld quality risk score, and residual stress risk index respectively.

10. A digital modeling and optimization system for the manufacturing process of large-span steel anchor boxes, characterized in that, include: The manufacturing process data package generation module collects geometric point cloud data, welding process parameters, environmental parameters, and equipment status parameters corresponding to the process nodes during the manufacturing process of the steel anchor box, and aligns and associates them based on a unified timestamp and component identifier to generate a manufacturing process data package. The steel anchor box manufacturing process knowledge graph establishment module establishes a steel anchor box manufacturing process knowledge graph based on component segmentation, assembly benchmark, weld type and sequence dependency relationship, and maps the manufacturing process data package to the attribute set of knowledge graph nodes and edges to form a queryable process semantic model. The convolutional neural network prediction model calculation module extracts the geometric semantic features, welding thermal input features, assembly constraint features and fixture features of the steel anchor box from the manufacturing process data package, and inputs them into the trained convolutional neural network prediction model, outputting the predicted value of the deformation of the key section after welding, the quality risk score of the key weld and the residual stress risk index. The risk penalty term generation module evaluates the uncertainty of the prediction model output to obtain the prediction confidence interval, and generates a risk penalty term based on the prediction confidence interval. The executable instruction set output and execution module, under the conditions of meeting equipment capacity, tooling and fixture accessibility, welding specifications, cycle time and quality constraints, performs multi-objective optimization on material cutting compensation, segmented assembly sequence, welding sequence, welding parameters, welding car path and fixture layout based on risk penalty items, and outputs and executes the executable instruction set; The recording module links the executable instruction set and manufacturing process data package to the knowledge graph and forms a traceable data chain.

Images