A discrete plant reconfiguration digital twin implementation method, system, device and medium

Abstract

The application discloses a discrete workshop reconstruction digital twin implementation method, system, equipment and medium, relates to the technical field of manufacturing management, and comprises the following steps: collecting multi-modal data of a workshop, and constructing a workshop layout image dataset; an initial instance segmentation is performed on the workshop layout image by using a Mask-RCNN model to obtain an original segmentation result; the original segmentation result is corrected based on semantic rules in an ontology knowledge base, including multi-dimensional verification of categories, boundaries and completeness; the corrected segmentation result is semantically mapped with the ontology knowledge to establish semantic association of equipment process layout; multi-scene simulation is performed on a candidate layout scheme in a digital twin environment, ontology reasoning rules are dynamically updated according to simulation errors, segmentation correction and layout decision are optimized, and an optimal reconstruction scheme is output and fed back to a physical workshop for layout adjustment. The application realizes simulation-driven updating and reasoning rule optimization of ontology knowledge, excavates the relationship between target objects, improves the instance segmentation accuracy and the matching degree of the reconstruction scheme.

Description

Technical Field

[0001] This invention relates to the technical field of manufacturing management, and in particular to a method, system, equipment and medium for realizing digital twins in discrete workshop reconfiguration. Background Technology

[0002] The advancement of new-generation information technology (big data, artificial intelligence, digital twins) and manufacturing technology toward intelligentization has brought about a tremendous transformation in global manufacturing. However, this transformation is underpinned by the upgrading of technological methods in the discrete manufacturing industry. In this process, the discrete manufacturing industry has encountered problems such as order fragmentation, personalized demand, and dynamic production.

[0003] Discrete manufacturing refers to a series of discontinuous production processes, including discrete processing, forming, and assembly. Its production process is characterized by discontinuity, process coupling, and dynamic resources. Different processing equipment, process parameters, and production resources are required. Furthermore, its production model accelerates product variety changes, resulting in a high-variety, low-batch production model. Temporary order insertions and batch changes during production and assembly can affect production planning. Because it involves the management of suppliers, logistics, production, and outsourced parts throughout the entire process, delays in information transmission and inadequate coordination are common, creating numerous "information silos" between different stages, hindering information sharing and collaboration, and preventing the dynamic optimization of production factors.

[0004] In discrete workshop layout analysis, there is a heterogeneity problem between low-level features of visual images (contours, textures), high-level semantic knowledge of the ontology (process rules, equipment associations), and dynamic data of physical equipment (operating status, logistics parameters). Especially in discrete workshops under dynamic reconstruction scenarios, as orders change frequently with the production scenario (temporary order insertion or large-scale adjustment), the inference rules in the static ontology model cannot adapt to this change, and "rule failure" is likely to occur, which will cause the reconstruction digital twin solution to fail to achieve the expected results. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a method, system, equipment, and medium for realizing digital twins of discrete workshop reconstruction, thereby solving the problems in the prior art.

[0006] This invention specifically provides the following technical solution: a method for implementing digital twins in discrete workshop reconstruction, comprising the following steps: Collect multimodal data in the workshop, including the location of each piece of equipment, the flow path of materials, and the operating status of the equipment, and construct a workshop layout image dataset; Initial instance segmentation is performed on the workshop layout image to extract the original segmentation mask, original category label and visual features of the equipment, and to construct the original segmentation results of each piece of equipment in the workshop. Construct and load an ontology knowledge base that stores equipment attribute knowledge, spatial relationship knowledge, and process constraint knowledge, and obtain the matching degree between the original segmentation result and the ontology knowledge base. If it is lower than the threshold, trigger the correction process, use ontology reasoning rules to correct the original segmentation result, and generate the corrected segmentation result. The corrected segmentation results are semantically mapped to the ontology knowledge base to establish semantic associations for equipment process layout, and candidate layout schemes are proposed based on the semantic information in the semantic associations. In a digital twin environment, candidate layout schemes are simulated in multiple scenarios. The ontology reasoning rules are dynamically updated based on the simulation error to optimize segmentation correction and layout decisions. The optimal reconstruction scheme is then output and fed back to the physical workshop for layout adjustment.

[0007] Preferably, if the matching degree between the original segmentation result and the ontology knowledge base is lower than a threshold, a correction process is triggered to correct the original segmentation result using ontology reasoning rules, generating a corrected segmentation result. Specifically: Based on the target category and location in the original segmentation results, the corresponding equipment attribute knowledge, spatial relationship knowledge and process constraint knowledge are retrieved from the ontology knowledge base through the retrieval function; Based on ontology reasoning rules, the original segmentation results are verified in multiple dimensions. If the results are below the threshold, a correction process is triggered. Ontology reasoning rules are used to correct the category misjudgment, boundary ambiguity, and target omission in the original segmentation results, and a corrected segmentation result is generated.

[0008] Preferably, the original segmentation result is verified in multiple dimensions according to ontology reasoning rules. If the result is below a threshold, a correction process is triggered. The ontology reasoning rules are used to correct category misclassification, boundary ambiguity, and target omission in the original segmentation result. Specifically: If the target category does not match the device attribute knowledge, it is corrected to a revised target category label based on the device structural characteristics. ; If the original segmentation mask If the corresponding device boundaries do not conform to spatial relationship knowledge, then adjust the original segmentation mask to the corrected segmentation mask. ; If small targets are missed, supplement them. to ,in For the added new target, For the boundary information of the new target, This provides category information for new targets; small targets are those with a pixel area smaller than a threshold.

[0009] Preferably, the step of dynamically updating the ontology reasoning rules based on simulation errors and optimizing segmentation correction and layout decisions specifically includes: By combining simulation errors, the ontology inference rules are optimized using an ontology rule update function, and then converted into physical entity adjustment instructions through an inverse mapping function to obtain a dynamic optimization scheme for the workshop layout.

[0010] Preferably, the simulation error is specifically: Simulations are performed on the potential layout schemes within a digital twin environment. A performance metric set, including production efficiency (Eff), logistics cost (Cost), and flexibility (Flex), is obtained using a performance index calculation function. P ; like P If the optimization objective is not achieved, a new solution is generated by iteratively updating the function, and the simulation is repeated until the optimal solution is obtained. t The simulation error is obtained by comparing the simulation output at any given time with the measured data from the physical workshop at the same time.

[0011] Preferably, the initial instance segmentation of the workshop layout image, extracting the original segmentation mask, original category label, and visual features of the equipment, and constructing the original segmentation results of each piece of equipment in the workshop, specifically involves: Visual features of the workshop layout image are extracted using ResNet in the Mask-RCNN model, and visual features at different levels are fused to obtain a fused feature map. On the fused feature map, set the anchor box scale set and aspect ratio set, and perform classification and regression processing to obtain the category score and candidate box coordinates. Then, filter overlapping candidate boxes to obtain the candidate region set. For the candidate region set, feature blocks are extracted using RoI Align through bilinear interpolation, and pixel-level segmentation of workshop equipment is performed using a dual-branch head to obtain the original segmentation results of each piece of equipment in the workshop; the original segmentation results include the original segmentation mask and the original category label.

[0012] This invention provides a digital twin implementation system for discrete workshop reconfiguration, comprising: The data acquisition module is used to collect multimodal data in the workshop, including the location of each piece of equipment, the flow path of materials, and the operating status of the equipment, and to build a workshop layout image dataset. The segmentation module is used to perform initial instance segmentation on the workshop layout image, extract the original segmentation mask, original category label and visual features of the equipment, and construct the original segmentation results of each piece of equipment in the workshop. The correction module is used to build and load an ontology knowledge base that stores equipment attribute knowledge, spatial relationship knowledge, and process constraint knowledge, and obtain the matching degree between the original segmentation result and the ontology knowledge base. If it is lower than the threshold, the correction process is triggered to correct the original segmentation result using ontology reasoning rules and generate the corrected segmentation result. The mapping module is used to semantically map the corrected segmentation results with the ontology knowledge base, establish semantic associations for equipment process layout, and formulate candidate layout schemes based on the semantic information in the semantic associations. The adjustment module is used to perform multi-scenario simulations of candidate layout schemes in a digital twin environment, dynamically update the ontology inference rules based on simulation errors, optimize segmentation correction and layout decisions, output the optimal reconstruction scheme, and feed the optimal reconstruction scheme back to the physical workshop for layout adjustment.

[0013] The present invention provides a computer device, including a memory and a processor. The memory stores a program, and when the program is executed by the processor, the processor performs the steps of the above-described method for implementing a discrete workshop reconstruction digital twin.

[0014] The present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method for implementing a discrete workshop reconstructed digital twin.

[0015] Compared with the prior art, the present invention has the following significant advantages: This invention, based on the concepts of deep learning and dynamic ontology-twin closed loop, proposes a method for realizing digital twin reconstruction of discrete workshops. It segments workshop layout images in dynamic scenarios, corrects the original segmentation results to address issues such as misclassification, blurred boundaries, and missed detection of small targets inherent in traditional deep learning instance segmentation. Then, it combines an ontology knowledge model to imbue the corrected segmentation results with semantic information, establishing semantic associations for equipment and process layout. Furthermore, it proposes a cross-workshop knowledge transfer mechanism to achieve common knowledge reuse, dynamically updates ontology inference rules based on simulation errors, optimizes segmentation correction and layout decisions, and outputs the optimal reconstruction scheme. This realizes simulation-driven updating of ontology knowledge and optimization of inference rules, uncovers relationships between targets in dynamically changing scenarios, and improves instance segmentation accuracy and the matching degree of reconstruction schemes. Attached Figure Description

[0016] Figure 1 This is a diagram illustrating the implementation path of the discrete workshop reconstruction digital twin based on deep learning and ontology in this invention; Figure 2 This is a diagram of the Mask R-CNN neural network model in this invention; Figure 3 This is a schematic diagram of the working principle of the correction factor in this invention; Figure 4 This is the geometric modeling diagram of the elements in this invention; Figure 5 This is a diagram of the Protégé ontology modeling tool interface in this invention; Figure 6 This is a schematic diagram of the discrete workshop digital twin architecture in this invention; Figure 7 This is a flowchart of the layout scheme verification and optimization method in this invention; Figure 8 This is a workshop image dataset diagram from the present invention; Figure 9 This is a performance diagram of the backbone network in this invention; wherein Figure 9 (a) is the average detection accuracy graph. Figure 9 (b) is the reasoning speed diagram. Figure 9 (c) is the segmentation average accuracy map. Figure 9 (d) is the model parameter diagram; Figure 10 This is a diagram of the ResNet-50 network structure in this invention; Figure 11 This is a graph showing the average accuracy versus the loss function in this invention; wherein, Figure 11 (a) is the loss function graph. Figure 11 (b) is the average accuracy plot; Figure 12 This is a model diagram of the ball bearing processing and assembly workshop in this invention; Figure 13 This is a flowchart of the discrete workshop layout image analysis system in this invention; Figure 14 This is a flowchart of a method for implementing a discrete workshop reconstructed digital twin, provided by the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0018] In existing technologies, there remains a challenge in the fusion and representation of heterogeneous data (visual images, semantic knowledge, and physical data). In discrete workshop layout analysis, the low-level features of visual images (contours, textures), the high-level semantic knowledge of the ontology (process rules, equipment associations), and the dynamic data of physical equipment (operating status, logistics parameters) exhibit heterogeneity. Current methods cannot effectively unify the representation of these three parts, resulting in segmentation results lacking semantic interpretability and failing to directly generate a reconstruction scheme that meets actual production needs. Based on this problem and other related issues, this invention discloses an optimized solution for discrete workshop reconstruction digital twin implementation methods, systems, equipment, and media. The system architecture uses "data-driven - segmentation modeling - ontology correction - knowledge mapping - twin verification - feedback optimization" as its core logic, and collaboratively completes workshop layout image analysis, semantic knowledge extraction, dynamic updates, and reconstruction scheme optimization in a layered manner. The details are described below.

[0019] Due to the challenges of diverse equipment types and complex logistics relationships in discrete workshop layouts, this invention addresses two main issues: heterogeneous data fusion and dynamic scenario adaptation. It proposes a digital twin implementation method for discrete workshop reconstruction based on deep learning and a dynamic ontology-twin closed loop. Using multi-source data as a foundation, and considering the multimodal data processing requirements of workshop scenarios, the physical and digital workshops are mapped one-to-one through the synergistic effect of the data layer, model layer, and twin layer. This completes the layout optimization from physical to digital workshops. Specific implementation steps and elements are as follows: Figure 1 As shown. Specifically, as... Figure 14 As shown, this embodiment provides a method for implementing digital twins in discrete workshop reconstruction, specifically including:

[0020] Step 1: Collect multimodal data in the workshop, including the location of each piece of equipment, the flow path of materials, and the operating status of the equipment, and construct a workshop layout image dataset.

[0021] The data layer (including processing areas, logistics areas, and assembly areas when designing the workshop layout) performs multimodal data acquisition and consistency assurance: The data layer obtains information such as the location of each piece of equipment in the workshop, the flow path of materials, and the operating status of the equipment. i Image data Iᵢ and the j Group point cloud data P j Finally, they are combined into a multimodal dataset. D ={ I 1, I 2,…, I j , P 1, P 2,…, Pj}, using data synchronization functions ⋅) This dataset D Synchronize to the model layer; the dataset received by the model layer is... D The point cloud data is processed using voxelization to extract three-dimensional geometric features. F cl It is used to extract industrial equipment attribute values ​​after semantic mapping.

[0022] To ensure consistency between virtual and real data, a consistency metric function is introduced. ⋅), its calculation formula is: (4); in For the physical layer raw data set, For the model layer to receive data sets, For the corresponding data sample, The mean of the Euclidean distance; fitted mapping function ( ),when hour, (Right now ), to meet the positioning error of workshop equipment The accuracy requirements.

[0023] The elements simply represent process information, layout planning, and resource information; among which, process information includes mechanical information, feature information, and technical requirements; layout planning includes typical cases, production line layout, overall design, design requirements, reasoning planning, reasoning rules, and design process; and resource information includes component information, component affiliation, and application information.

[0024] The model layer is described below, including instance segmentation, ontology correction, and knowledge mapping, specifically: Step 2: Use the Mask-RCNN model to perform initial instance segmentation on the workshop layout image, extract the original segmentation mask, original category label and visual features of the equipment, and construct the original segmentation results of each piece of equipment in the workshop.

[0025] Instance segmentation module: Uses the Mask-RCNN model to segment a subset of workshop layout images. ( Initial segmentation was performed on all image data (from the image), and visual features of the workshop layout image were extracted using a ResNet backbone network. Visual features from different levels are fused using FPN to obtain a fused feature map. RPN generates candidate regions, RoI Align extracts feature blocks, and finally outputs an initial set of instance segmentation results. A single raw result ,in For the first The original segmentation mask of each target. For the first The original category label for each target; specifically, such as Figure 2 As shown.

[0026] 1. Initial segmentation of Mask-RCNN: The backbone network and feature extraction (Backbone + FPN) are as follows: ResNet backbone: After inputting the workshop image, deep features are extracted through the residual blocks (11 in total) of ResNet. The formula is: (7); in The input is a workshop image; For the first ResNet The computation function for each residual block; This indicates that ResNet contains 11 residual blocks; This is the deep feature map output by the backbone network. Specifically, stage0, after undergoing 7×7 convolution, BN, ReLU, and max pooling, has an output size of [size missing]. The feature map is then used for further multi-scale feature extraction in subsequent stages.

[0027] FPN Feature Fusion: Fusion of feature maps from different levels, including high-level feature maps. (Experience the vastness of the field and capture global information from large-scale equipment) After 1×1 convolution dimensionality reduction, upsampling is performed and compared with low-level feature maps. (Rich in detail, preserving edge information of small tools) Adding them together yields a fused feature map. The formula is: (8); in This is a low-level feature map (from a shallow stage, rich in detail, used to preserve edge information of small tools). This is a high-level feature map (from the deep stage, sensing the vastness of the field, used to capture global information of large equipment). High-level feature map The result after dimensionality reduction using 1×1 convolution; For upsampling operation, make Size and match; This is the fused feature map, which preserves both details and global information.

[0028] 2. The Region Proposal Network (RPN) generates candidate regions: In the fusion feature map Above, set the anchor point frame scale set. and aspect ratio set (Based on the actual dimensions of the workshop equipment, statistical adaptation is performed), and classification and regression processing are carried out to obtain category scores and candidate box coordinates. Overlapping candidate boxes are filtered to obtain a set of candidate regions.

[0029] Among them, classification branches: through Differentiate between "device / background" and output category scores; regression branch: through The anchor box positions are corrected to obtain the candidate box coordinates; post-processing: overlapping candidate boxes are filtered using NMS (non-maximum suppression, threshold 0.7) to obtain the candidate region set. .

[0030] 3. RoI Align extracts candidate region features: For candidate region set RoI Align is used to extract feature blocks through bilinear interpolation. The formula is: (9); in The pixel values ​​are for 4 interpolation points. Weights are assigned to the distances from each interpolation point to the target sampling point. The feature values ​​of the sampling points are used to achieve precise alignment between the candidate region and the feature map.

[0031] 4. Dual-branch head (Box Head + Mask Head): Pixel-level segmentation of workshop equipment is performed using a dual-branch header to obtain the original segmentation results for each piece of equipment in the workshop. The original segmentation results include the original segmentation mask and the original category label. The following provides a detailed explanation of classification, bounding box regression, and mask generation:

[0032] Box Head: Yes After processing by fully connected layers (FC layers), the output class score (to determine the device class) and bounding box coordinates (to refine the target location) are obtained.

[0033] Mask Head: Generates the original mask through transposed convolution (deconvolution). After binarization with a threshold of 0.5, the original segmentation result is obtained. This completes pixel-level segmentation of the workshop equipment.

[0034] In summary, Mask R-CNN achieves accurate segmentation of equipment in workshop images through a process of "feature extraction - candidate region generation - feature alignment - classification / box regression / mask generation". It can accurately identify the equipment category and location, and finely segment the pixel region of the equipment.

[0035] Step 3: Construct and load an ontology knowledge base that stores equipment attribute knowledge, spatial relationship knowledge, and process constraint knowledge, and obtain the matching degree between the original segmentation result and the ontology knowledge base. If it is lower than the threshold, trigger the correction process, use ontology reasoning rules to correct the original segmentation result, and generate the corrected segmentation result.

[0036] Ontology reasoning correction module: Based on ontology concept set (e.g., processing unit) and relation set (e.g., "adjacent") construct semantic knowledge triples ( c 1, r , c 2) c 1, c 2∈ , r ∈ ), through the retrieval function ⋅) Retrieve knowledge such as process rules, equipment attributes, and spatial relationships from the ontology knowledge model (KG) to form a set. K ={ k 1, k 2,…, k}

[0037] Visual features and semantic features are matched using a specific mechanism: first, the original segmentation features are compared... Standardize to the [0,1] interval using Min-Max, the formula is as follows: Then, the ontology concepts are trained on a domain corpus to create 256-dimensional semantic vectors. Mapped to 1024 dimensions through a fully connected layer, resulting in... Dimensionally consistent semantic embedding features .

[0038] Will With K-input reasoning engine, based on ontology rules Calculate matching degree If there are issues such as incorrect equipment category labeling, spatial location conflicts, or unclear boundaries, a correction function will be triggered. Generate a single correction result ( For the corrected segmentation mask, (For the corrected category labels), the final set of corrected segmentation results is formed. .

[0039] During the revision process, a knowledge constraint term C is introduced, calculated as follows: (5); in, α =0.6、 β =0.4 (determined through 5-fold cross-validation). sim (⋅) represents the initial segmentation features and device attribute knowledge. (e.g., cosine similarity of "CNC machine tool machining accuracy ≥ 0.01mm") valid (⋅) indicates whether the original segmentation result satisfies the spatial relation knowledge. A Boolean function (e.g., 1 if satisfied, 0 if not satisfied) is used to guide the accurate correction of the segmentation results.

[0040] To improve the accuracy of corrections, an error correction mechanism is introduced for optimization, such as... Figure 3 As shown: Knowledge retrieval: based on the original segmentation results Target categories ,Location Information such as these, retrieved through a function From the ontology knowledge base Retrieve the corresponding device attribute knowledge Spatial Relationship Knowledge Process constraint knowledge .

[0041] Consistency check: Based on ontology reasoning rules, the original segmentation results are checked. Multi-dimensional verification is performed. If the result falls below a threshold, a correction process is triggered. Ontology reasoning rules are used to correct category misclassification, boundary ambiguity, and target omission in the original segmentation result, generating a corrected segmentation result. Specifically: Category validation: If the target category Device attribute knowledge If there is a mismatch (e.g., labeling "CNC machine tool" as "machining center"), then correct the target category label according to the equipment's structural characteristics. .

[0042] Boundary verification: If the original segmentation mask The corresponding equipment boundaries do not conform to spatial relationship knowledge (If the device spacing is <0.5m), then the original segmentation mask is converted to the corrected segmentation mask. .

[0043] Integrity check: If small targets are missed (e.g., detection instruments with pixel area <32×32), then supplement. to Small targets refer to targets whose pixel area is smaller than the threshold.

[0044] Output results: After verification and correction, the final segmentation result is output. .

[0045] A digital layout model is constructed based on the corrected segmentation results, and key indicators are analyzed, such as Figure 4 As shown: The definition of a CNC machine tool quadruple is... ( The coordinates of the bottom left corner of the base are: (Length and width); a cuboid shelf five-element group is defined as follows: ( The three-dimensional coordinates of the bottom left corner of the base are: (Length, width, and height); a cuboid conveyor belt structure is defined as... ( The coordinates of the starting point, For length, width, and height, (This refers to the direction vector).

[0046] Element geometric modeling: CNC machine tools: Calculation based on bounding box , Based on the origin of the workshop coordinate system (such as the lower left corner of the gate). The actual length is calculated by converting the pixel size (a mask width of 100 pixels corresponds to 5 meters).

[0047] Shelves: Expanded by combining layer information , This is the height of the basement (usually 0.5 meters). The total height (e.g., a 3-layer shelf is 2.4 meters high).

[0048] Conveyor belt: via direction vector (as along) (Positive direction of axis) definition Starting point This is the location of the conveyor belt entrance.

[0049] Spatial Relationship Analysis: Adjacency determination: If the machine tool With shelves satisfy and (If a 0.5-meter passage is reserved, which complies with the body's safety rules), then it is determined to be an adjacent layout.

[0050] Logistics route planning: based on equipment center coordinates Calculate the total path length from the raw material area to the machine tool area and then to the finished product area. ( (The distance between the two points).

[0051] Equipment utilization rate assessment: Based on production logs, equipment utilization rate is calculated using a formula: (10); in For equipment In time Usage status (0 for unused, 1 for used); This represents the total simulation time.

[0052] Flexibility index calculation: Based on order response speed and equipment adaptability, it is defined as: (11); in For the number of order types, For the first Response time for similar orders For the equipment to the first The degree of compatibility of order type (0~1).

[0053] Step 4: Perform semantic mapping between the corrected segmentation results and the ontology knowledge base to establish semantic associations for equipment process layout, and formulate candidate layout schemes based on the semantic information in the semantic associations.

[0054] Knowledge Mapping Module: Based on the corrected segmentation results With ontology knowledge K Establish a set of mapping relationships This allows the segmentation results to carry semantic information, determining equipment type, process matching degree, and logistics relevance, providing data support for the subsequent generation of reconstruction schemes. The ontology relation matrix is ​​decomposed into a 256×256 dimensional matrix. To conduct research on the quantitative correlation between concepts.

[0055] Step 5: Perform multi-scenario simulations of candidate layout schemes in the digital twin environment, dynamically update the ontology reasoning rules based on simulation errors, optimize segmentation correction and layout decisions, output the optimal reconstruction scheme, and feed the optimal reconstruction scheme back to the physical workshop for layout adjustment.

[0056] Twin Layer: Reconstruction Optimization and Closed-Loop Feedback Through the virtual-real mapping function ⋅) Establish physical workshop status With digital model state The association satisfies A set of potential layout schemes in a digital twin environment. The proposed scheme was simulated, and the performance index was calculated using a function. We obtain a performance set that includes production efficiency (Eff), logistics cost (Cost), and flexibility metrics (Flex). .

[0057] like P If the optimization goal is not achieved, update the function iteratively. Generate new solutions Repeat the simulation until the optimal solution is obtained. , where O( slay (Based on weight optimization) α , β , γ The target function constructed (corresponding to Eff, Cost, Flex).

[0058] To achieve dynamic optimization, the first... k The rule confidence W(t) of the ontology rule at time t, and the actual error of the digital twin simulation guided by the rule. Its quantification formula is: (6); in for t Real-time simulation output (such as equipment utilization). These are measured data from the physical workshop during the same period. T The simulation duration is used. Simulation errors are substituted into the dynamic rule update formula to optimize the ontology reasoning rules, achieving a closed-loop cycle of "segmentation-correction-mapping-simulation-rule update-re-correction".

[0059] The ontology is constructed using a hierarchical modeling approach of "concept-attribute-relationship-instance-dynamic update," and the specific method is as follows: like Figure 5 The diagram shows all the core elements of ontology concept and attribute modeling, modularly divided into five parts: process information, resource information, layout decisions, cases and data, and knowledge interaction. It also constructs the knowledge flow process between modules and determines the core and derived concepts of each module. Concept instances are created through instance construction and verification, and their attributes are associated. Logical consistency is then tested using an inference engine. (Inference rule set) Based on the SWRL language definition, it covers the rules for segmentation result correction and the core rules for reconstruction decisions. During the rule reasoning process, the inference engine follows... right middle Analyze each instruction individually, generate correction instructions, and output them. At the same time, combined with digital twin simulation error Updated using formula (2) This enables dynamic optimization of rule confidence. The specific reasoning rules for the workshop system ontology knowledge model are as follows: 1. Rule: If equipment ?x is installed in region ?y (be installedin If (x, y) is a given condition, then region ?y must contain the device (comprise(?y, ?x)). The SWRL expression for this rule is: beinstalledin(?x, ?y) → compute(?y, ?x); 2. Rule: If the linear conveyor ?b receives input material ?x, then ?b will necessarily output material ?x (output material properties are consistent with input). The SWRL expression for this rule is: LinearConveyor(?lc)∧Material(?inMat)∧receivesInputMaterial(?lc, ?inMat) → Material(?outMat)∧outputsMaterial(?lc, ?outMat)∧hasPropertiesMatch(?inMat, ?outMat); 3. Rule: If a warehouse shelf ? exists, and the shelf is connected to the warehouse, then the warehouse must be connected to the warehouse. If the material output is to the conveyor, then the conveyor must be a pallet conveyor. conveyorbelt The SWRL expression for the rule is: Warehouse Shelf (??shelf)∧belt outputs (??shelf,?conveyor)→Pallet conveyorbelt (conveyor); 4. Rule: If the machining accuracy of the CNC machine tool ?m is ≥0.01mm, and the accuracy requirement of the workpiece ?w is ≤0.01mm, then ?m can machine ?w. The SWRL expression of the rule is: CNC Machine(??m)∧Workpiece(??w)∧hasPrecision(??m,?p1)∧requiresPrecision(??w,?p2)∧swrlb:lessThanOrEqual(?p2,?p1)→canProcess(??m,?w); 5. Rule: If device ?e1 and device ?e2 are upstream and downstream, and the distance between them is <0.5m, then the boundaries of the two overlap or the distance is unreasonable in the original segmentation result, and the segmentation mask needs to be corrected. The SWRL expression of the rule is: hasUpDownRelation(??e1,?e2)∧hasD istance(??e1,?e2,?d)∧swrlb:lessThan(?d,0.5)→needCorrectMask(??e1,?e2); 6. Rule: If the number of AGVs is ≥4 and the logistics efficiency improvement rate is <5%, then the number of AGVs does not need to be increased. The SWRL expression of the rule is: AGV(??a)∧hasQuantity(??a,?n)∧swrlb:greaterThanOrEqual(?n,4)∧efficiencyImprovement(??a,?r)∧swrlb:lessThan(?r,0.5).05) → noNeedAdd(??a); 7. Rule: If the initial segmentation label category is "machining center", but the equipment structure features match the "CNC machine tool" attribute, then correct the category label to "CNC machine tool". The SWRL expression for the rule is: $\mathrm{SegmentResult}(?s)\land\mathrm{rawCategory}(?s,"\mathrm{\text{machining center}}")\land\mathrm{hasStructure}(?s,?str)\land\mathrm{cncStructure}(?str)\rightarrow\mathrm{correctCategory}(?s,"\mathrm{\text{CNC machine tool}}")$ 8. Rule: If a small detection instrument (pixel area < 32×32) does not appear in the original segmentation result, and the workshop layout plan requires the detection instrument to be placed at that location, then supplement the segmentation mask and category label. The SWRL expression of the rule is: $\mathrm{WorkshopLayout}(?l)\land\mathrm{requiredEquipment}(?l,"\mathrm{\text{Detection Instrument}}")\land\mathrm{position}(?l,?pos)\land\mathrm{noSegmentResult}(?pos)\land\mathrm{smallEquipment}("\mathrm{\text{Detection Instrument}}")\rightarrow \mathrm{addSegmentResult}(?pos,"\mathrm{\text{Detection Instrument}}")$.

[0060] Taking rules as an example, a formalized description method is used to present them using first-order logic, ensuring the rigor of the reasoning. Based on Results of the initial segmentation The analysis is performed, and correction instructions are generated and output during the reasoning process. The confidence level of the rules is dynamically updated based on the results of digital twin simulation during the reasoning process.

[0061] Layered ontology model: Construct a layered structure of "common workshop knowledge + personalized knowledge". Common knowledge (such as "equipment safety distance ≥ 0.5m", "machining accuracy matching rules", and "general rules for segmentation result correction") is shared by all discrete workshops, while personalized knowledge (such as "quenching tank temperature constraints" and "special equipment segmentation correction rules" for the bearing workshop) is customized for specific workshops.

[0062] The knowledge transfer function uses cosine similarity to filter common knowledge between the source and target workshops. The formula is: (12); in For the source workshop ontology concept set, Let the set of ontological concepts of the target workshop be the numerator, the number of common concepts between the two workshops be the denominator, and the total number of concepts between the two workshops be the denominator (to avoid double counting), with a value range of [0,1]. When a workshop is identified as having high similarity, a migration strategy of "common knowledge reuse + personalized knowledge supplementation" is adopted.

[0063] Process similarity calculation: (13); in , These are the process sets for the source workshop and the target workshop, respectively. If necessary, perform a full migration; otherwise, only migrate common process knowledge.

[0064] Digital twin-ontology closed-loop architecture implementation method as follows Figure 6 As shown: This framework is based on the actual physical entities of the workshop. Based on, using device space coordinates To determine the location of the device, and using the device attribute feature vector The system represents the processing accuracy, rated power, and other attributes of the equipment, mapping the original workshop and the renovated workshop into basic digital representations; it also configures an IoT sensor network to collect equipment operating status, logistics parameters, and IoT feedback data, utilizing existing historical datasets in WDDS. and real-time acquired data sequences ( Data resources shared in public folders (including timestamps) and public folders are used to obtain a standard dataset through data fusion methods. (Including image feature data, equipment status data, logistics data, etc.), providing high-quality multi-source heterogeneous data for model training and simulation. A parametric modeling approach is adopted: first, virtual initial parameters are defined. Then through the conversion function Optimize the initial parameters, and then map them using the virtual model function. Will Convert to virtual model parameter vector For geometric parameters, For kinematic parameters, (For dynamic parameters), by modifying the virtual model parameter vector This achieves accurate reproduction of physical entities while updating the system's rule base. It utilizes digital twin technology to build simulation scenarios and employs multi-index evaluation functions. ( As weight, (A single indicator function) combined with dynamically adjustable The system comprehensively evaluates indicators such as equipment utilization, logistics efficiency, and flexibility in this scenario, and uses the simulation results as input to a dynamic ontology. The ontology inference rules are then optimized using a rule update function. This is done from a set of layout schemes. Obtain the virtual parameter values ​​corresponding to the optimal layout scheme and apply the feedback control function. ( For optimal parameters, For the current parameter, (For adjustment purposes) The virtual parameter values ​​corresponding to the optimal layout scheme are converted into physical entity adjustment instructions. It provides layout guidance to the renovated workshop and completes the decision optimization and execution feedback loop based on this; it stores the relevant knowledge of the optimal solution in the ontology knowledge base (PK) and public folder for use in subsequent workshop reconstruction.

[0065] The layout scheme verification and optimization follows a closed-loop process of "virtual modeling - simulation evaluation - rule update - scheme iteration - physical mapping", with specific steps as follows: Figure 7 As shown: Initial layout modeling: based on the results corrected by ontology reasoning. and the obtained ontology reasoning relationship and process constraints In the device component library Select the corresponding device, invoke the cross-workshop knowledge transfer mechanism to initialize the ontology model, and call the layout planning function. Obtain the initial virtual layout model .

[0066] Multi-scenario simulation experiments: This section sets up four scenarios: peak order period, equipment maintenance, process switching, and dynamic switching of multiple orders. Each scenario is simulated for 72 hours, and factors such as logistics path, equipment status, and process connection are monitored in real time. In order to ensure the randomness of the experiment, each scenario is simulated 5 times, and the average value of the indicators is taken as the result each time.

[0067] Scheme evaluation and parameter optimization: Using simulation data Substitute it into the quantization scheme, and according to the optimization algorithm Solve for the optimal parameters Using iterative update function Optimized Updated virtual model Using the differences in indicators between the alternatives as the comparison object, statistical significance (t-test, significance level) is used to determine whether there are sufficiently large differences between the alternatives to achieve the optimization effect. Significance test ( t Test, significance level Analyze the differences in indicators between different schemes to ensure the reliability of the optimization effect.

[0068] Ontology rule update: incorporating simulation errors Update function using ontology rules Optimize ontology reasoning rules to improve their robustness and segmentation correction accuracy.

[0069] Virtual-Real Mapping and Physical Adjustment: Through Inverse Mapping Functions Will Convert into physical entity adjustment instructions This process yields a dynamic optimization scheme for the workshop layout, and repeats until the desired multi-objective optimization scheme is obtained.

[0070] The ball bearing processing and assembly workshop was selected as the experimental object, and the experimental process followed the framework of "data-driven - segmentation modeling - ontology correction - knowledge mapping - simulation verification - closed-loop optimization".

[0071] In the selection of backbone networks for Mask R-CNN, ResNe50 / 80 / 101 and other backbone networks can be selected. The following describes the application practice of each backbone network in workshop layout instance segmentation and the parameter comparison, in conjunction with the specific process and experimental design.

[0072] 1. Dataset Construction: Aligned with Real-World Workshop Scenarios: (1) Data collection: through methods such as Figure 8 As shown, this design selects the reconstruction of the ball bearing processing and assembly workshop as the main research object, and uses a workshop layout image acquisition method based on digital twin simulation to collect a total of 2,000 workshop layout images, including two different lighting environments (natural daylight and artificial nightlight) and two different shooting angles (overhead and side shots), which are divided into 10 categories.

[0073] Labeling and partitioning: The workshop components were labeled using the Labelme image annotation tool. The dataset was in COCO format and divided into training, validation and test sets in an 8:1:1 ratio.

[0074] Image preprocessing: This experiment addresses issues such as uneven lighting, equipment occlusion, and complex backgrounds in discrete manufacturing workshop layout images through preprocessing.

[0075] Image scaling: Based on the actual workshop equipment and image resolution (1 pixel = 0.05 meters, corrected according to actual workshop measurement data), the image is scaled while maintaining the relative positional relationship between the equipment.

[0076] Contrast enhancement: Improving local contrast can improve the unclear features on the metal surface of a machine tool caused by reflection.

[0077] Data augmentation: The following random augmentation methods were used to increase the amount of data to prevent model overfitting: random rotation (-15°~15°), random horizontal flip (probability of 0.5), Gaussian noise (variance = 0.01), and brightness modification (±10%).

[0078] 2. Model Training and Parameter Setting: A Mask-RCNN model based on the PyTorch framework (with ResNet as the backbone) was tested on a Windows 11 operating system, using CUDA 11.3 and Python 3.8. The initial learning rate was set to 0.0001, weight decay to 0.0005, momentum to 0.9, batch size to 4, and 100 training epochs. AdamW was chosen as the optimizer, and cosine annealing was selected as the learning rate decay strategy. , An early stopping strategy is adopted: if the validation set mAP does not improve within 5 consecutive rounds, the current training ends. A grid search for the hyperparameters was performed based on the above hyperparameters to find the optimal solution, thus obtaining the hyperparameters for this configuration.

[0079] Depend on Figure 9 As shown, the detection and segmentation mAP increases progressively with model depth from ResNet50 to ResNet80 and ResNet101, but the gain gradually decreases, reaching 76.3, 82.3, and 83.1 respectively. Deeper models have longer inference times (32 and 15 respectively) and significantly increased parameter counts (100M and 500M respectively), resulting in high computational costs and resource consumption. Considering both performance and efficiency in the workshop layout instance segmentation scenario, ResNet50 is the optimal choice for balancing performance and efficiency in this scenario, and it also exhibits good adaptability. Figure 10 The diagram shows the network structure of ResNet50.

[0080] ResNet-50's shallow features can capture details such as device edges, screw holes, and conveyor belt textures, enabling accurate segmentation of small objects, such as small tool racks. Deep features provide information like the overall machine tool structure and shelf hierarchy, preventing the "machine tool base" from being mistaken for the "ground." Compared to ResNet80, it offers higher semantic discriminative power, fewer parameters than ResNet101, and faster speed, meeting real-time requirements. Based on mature open-source pre-trained models, transfer learning allows for rapid adaptation to workshop datasets at a lower cost. Furthermore, by optimizing shallow convolutions, it can adapt to the color characteristics of industrial scenarios and the complex environment of workshops.

[0081] like Figure 11 As shown, during the training process of the ResNet50-MaskRCNN model, the mAP remained relatively stable around 0.8 without significant overfitting. Instance segmentation performed well, and the accuracy of bounding box localization and mask generation was relatively high. The initial Training Loss value was high, then dropped sharply in the initial stage before stabilizing. The prediction error during training gradually decreased and stabilized with increasing training iterations, indicating that the model's training performance improved overall. The learning rate (lr) fluctuated initially before stabilizing at a low level. This value represents the step size used by the learner to update parameters. A larger learning rate initially facilitated rapid convergence, while a smaller learning rate was used to fine-tune the model once stability was achieved. The results show that the model converged quickly in the early stages and entered a stable learning state later. The training process was relatively reasonable and stable.

[0082] like Figure 12 As shown, the seven-step method is used to construct the ontology. First, the main objects are identified, such as the bearing assembly area, the bearing inner ring, the bearing outer ring, its blank (bearing ring processing area), and the finished product transportation area. Their connotations are defined in the form of classes, and their size and materials are described by properties. The relationships between them are expressed by object properties, thereby realizing the physical entities as the basic material support for important components of workshop production, that is, as the underlying knowledge elements in the ontology model.

[0083] Based on the ontology mapping method, abstract process operations such as ConditionaIOperation, DelayOperation, and TransportOperation are explicitly defined at the process level and classified into process subclasses. Specific processing techniques, such as the process of bearing ring blanks undergoing heat treatment in FurnaceOvens, welding in an ERSX Reflow Soldering oven, and finishing in a Machining Center, are defined using data properties to determine the sequence, equipment, objects, and process parameters (e.g., processing temperature, time thresholds) of each process. Furthermore, Semantic Web Rule Language (SWRL) is used to define rules, clearly indicating the semantic relationships between production processes.

[0084] The rule system determines the logical consistency and accuracy of the ontology model, and descriptive logic is used for standardized definitions. (x) constrains material movement from the entity and process definition perspectives (e.g., the TransportOperation execution rule for expressing the movement of bearing ring blanks from the storage area to the processing area: expressed as SWRL rule: BearingRingBlank(x)∧StorageArea(y)∧ProcessingArea(z)Transport(x,y,z)). The process is constrained by the execution rules of the process operation: in ConditionalOperation, it is stipulated that the material's size, temperature, and other data must meet certain standards before the subsequent process can continue; expressed as SWRL rule: Material(m)∧HasDimension(m,d)∧DimensionThreshold(t)∧swrlb:greaterThanOrEqual(d,t)NextProcess(m). There are also operation rules for equipment: furnaces, reflow ovens, and other equipment have strict requirements for parameters such as temperature and time. This effectively constrains the state changes during equipment operation. The rule expression is: Equipment(?e)∧Operation(?o)∧HasTemperature(?e,?temp)∧TemperatureRange(?r)∧swrlb:inside(?temp,?r)ValidOperation(?o). Secondly, an association between data and visualization is established through ontology data mapping. Using a resource description framework (RDF), physical entity attribute data (size, material) and process parameter data (processing time, temperature) are obtained. Ontology models are used for data fusion and semantic enhancement. Furthermore, a visualization factory platform is utilized to directly reflect the workshop's processing status through visualization methods such as flowcharts and equipment operation status diagrams driven by web ontology language (OWL), providing support for workshop production management and optimization decisions, and promoting the transformation of the ball bearing processing and assembly workshop towards an intelligent manufacturing model.

[0085] A workshop layout image analysis system was developed based on PyCharm, integrating Mask-RCNN instance segmentation, ontology reasoning correction, and workshop knowledge retrieval. It combines objective functions to perform dynamic ontology modeling, reasoning decision-making, and digital twin simulation, realizing the entire process analysis from image segmentation to correction, knowledge matching, and scheme optimization.

[0086] The workflow of the discrete workshop layout image analysis system includes three main steps: input element loading, image analysis and information extraction, and result display and analysis. Figure 13 The entire process is linked together to form a complete process of image analysis and information extraction, and the entire process can be directly displayed on the system's operation interface.

[0087] Input element loading: (1) Model loading: According to the specified model file path, combine the label mapping file to establish the corresponding mapping between the model output and the workshop element category, and set the confidence threshold (e.g., 0.5); (2) Ontology file loading: Open the ontology knowledge base of the workshop domain; (3) Image loading: Import the workshop layout diagram according to the address of the given diagram to be analyzed.

[0088] 2. Image analysis and information extraction: The Mask-RCNN model is called to identify workshop elements and generate segmentation results. Then, the segmentation results are matched with ontology concepts in the ontology knowledge base to infer ontology concepts for semantic mining and expansion of semantic information.

[0089] 3. Results Presentation and Analysis: Figure 13 The results display area on the right presents the specific size of the corresponding equipment, path or functional area as a "Processed Image"; the table presents specific information including element ID, category, confidence level, area (pixels), attributes, etc.; and displays detailed descriptions of the elements (function and composition of "Straight Conveyor") and accuracy (precision, recall, etc.) for use during shop floor reconstruction. Figure 13 The document showcases units for Mask-RCNN, ontology file, image to be processed, result comparison, result display, image processing, deep learning, HermiT, digital twin, information matching, confidence display, and detailed information; these are obtained through image selection, image processing, and result analysis.

[0090] Two sets of experiments were set up: "without ontology reasoning (initial segmentation only with Mask-RCNN)" and "with ontology reasoning correction (Mask-RCNN + ontology reasoning)". Segmentation performance and reconstruction results were compared. The experiments show that the segmentation performance comparison between "without ontology reasoning" and "with ontology reasoning correction" is shown in Table 1. After ontology reasoning correction, the segmentation mAP reached 93.5%, which is 8.2% higher than the traditional Mask-RCNN without ontology reasoning; the segmentation accuracy for small objects was also improved. Improved the segmentation accuracy of occluded targets by 7.8%. The accuracy improved by 8.6%, precision by 6.3%, and recall by 7.5%, verifying that the ontology reasoning correction mechanism has a good corrective effect on segmentation errors. mAPocc

[0091] Table 1. Comparison of Segmentation Accuracy Three sets of control experiments were set up: "No ontology guidance (Mask-RCNN only)", "Static ontology guidance (Mask-RCNN + static ontology correction)" and "Dynamic ontology - Siamese loop closure (Mask-RCNN + dynamic ontology correction + loop closure feedback)" to verify the improvement effect of each module on the reconstruction effect.

[0092] The "Dynamic Ontology-Digital Twin Closed Loop" group outperformed the other two groups in all indicators. The production efficiency was 18.5% higher than that of the "Unintegrated Ontology Reasoning" group, the equipment utilization balance was improved by 32%, the flexibility index was improved by 25%, and the reconstruction scheme generation time was shortened by 33.3%. This verified the effectiveness of the dynamic ontology and digital twin closed loop mechanism, as shown in Table 2.

[0093] Table 2 Performance Comparison of Refactoring Schemes with Different Module Combinations To verify the applicability of the method of this invention to dynamic scenarios, three order scenarios were designed: small-batch multi-variety (100 pieces of each of 5 varieties), large-batch few-variety (500 pieces of each of 2 varieties), and temporary insertion (300 urgent workpieces inserted during batch production). Simulation verification was conducted, and the results are shown in Table 7. To further ensure the rationality of the simulation results, five simulation experiments were conducted in each scenario. The results of each simulation were identical; therefore, only the average value of each simulation experiment is given here. The coefficient of variation (CV) is used to measure the stability of the data. For the simulation results of a certain algorithm, the CV must meet the requirement of <10% to ensure the repeatability of the experiment.

[0094] As shown in Table 3, the proposed method reduces the reconstruction scheme generation time by 32%-36% compared to the "without ontology reasoning" group during scene switching, with the largest saving (36.8%) for temporary order insertion scenarios. This is because using a dynamic ontology allows the invocation of historical order insertion processing rules, eliminating the need for reconstructing the reasoning logic. Equipment utilization fluctuations are reduced by 21%-25%, with the best fluctuation control for small-batch, multi-variety scenarios (from 18.5% to 14.6%), demonstrating the closed-loop system's good ability to balance multi-task loads. Flexibility is improved by 16%-19%, with the largest increase in flexibility for large-batch, few-variety scenarios (from 79.5% to 91.2%), indicating good adaptability of the dynamic ontology to equipment. Order response time is reduced by 22%-27%, with the largest improvement in response efficiency for urgent order insertion scenarios (from 10.2 h to 7.5 h), demonstrating that this method effectively alleviates production uncertainty. All indicators have coefficients of variation less than 8%.

[0095] Table 3 Performance Comparison of Multi-Order Dynamic Switching Scenarios (Mean ± Standard Deviation) The automotive engine assembly workshop was selected as the target workshop, and the process similarity between the two workshops was calculated. (Judged as high similarity). Among them (The bearing workshop) includes 12 core processes. The engine workshop contains 15 core processes, with 9 overlapping processes (precision machining, heat treatment, etc.), employing a complete knowledge transfer strategy.

[0096] To clarify the core value of knowledge transfer, a specific comparison of the ontological knowledge systems of the two workshops is first conducted to identify the core content of reuse and supplementation, as shown in Table 4: Table 4. Cross-workshop knowledge transfer comparison table Comparing the effects of "no migration (building the ontology from scratch)" and "with migration (transferring knowledge from the bearing workshop)": the ontology construction time for the migration group was reduced by 62% (from 15 days to 5.7 days), and the amount of manual annotation was reduced by 40% (from 1000 sheets to 600 sheets); the productivity improvement was significant. , and the non-migrating group ( The difference was less than 3% (p=0.42>0.05), proving that the migration mechanism reduces costs while ensuring the reconstruction effect.

[0097] Table 5 Comparison of cross-workshop knowledge transfer effects (mean ± standard deviation) Note: A p-value > 0.05 indicates that there is no statistically significant difference between the two groups, proving the effectiveness of the transfer program.

[0098] The relevant model functions are as follows: 1. Segmentation results and ontology semantic matching model: (1); in For the first The original segmentation results, For the first Ontological knowledge, The original segmentation visual feature vector, It is a 256×256 dimensional ontology relation matrix. For a 256-dimensional ontology concept semantic vector; when If the segmentation result does not match the ontology knowledge, a correction process is triggered.

[0099] 2. Dynamic rule update model: (2); in For the first Rule in Confidence level at any given moment Update the learning rate for the rules (by...) Comparative experiments verified that this value achieved a mAP of 93.5% and had the shortest reconstruction time. To allow for the maximum simulation error, For the first The real-time simulation error of the rule; when hour, Increasing the number raises the rule priority.

[0100] 3. Multi-objective optimization function: (3); in ( (a set of layout schemes) Production efficiency ( Logistics costs ), flexible indicators ( The weights of ) satisfy and Retrieving data in dynamic scenarios with multiple orders (Determined through the Analytic Hierarchy Process) "Negative cost terms" means that cost reduction will increase the objective function value, which is consistent with the direction of efficiency and flexibility optimization.

[0101] Based on the above method, the present invention also provides a discrete workshop reconstructed digital twin implementation system, comprising: The system comprises the following modules: a data acquisition module for collecting multimodal data on the location of each piece of equipment, the flow path of materials, and the operating status of equipment within the workshop; a segmentation module for initial instance segmentation of the workshop layout image, extracting the original segmentation mask, original category label, and visual features of the equipment, and constructing the original segmentation results for each piece of equipment in the workshop; a correction module for constructing and loading an ontology knowledge base that stores equipment attribute knowledge, spatial relationship knowledge, and process constraint knowledge, and obtaining the matching degree between the original segmentation results and the ontology knowledge base. If the matching degree is lower than a threshold, a correction process is triggered, using ontology inference rules to correct the original segmentation results and generate corrected segmentation results; a mapping module for semantic mapping between the corrected segmentation results and the ontology knowledge base, establishing semantic associations for equipment process layout, and proposing candidate layout schemes based on the semantic information in the semantic associations; and an adjustment module for performing multi-scenario simulations of candidate layout schemes in a digital twin environment, dynamically updating ontology inference rules based on simulation errors, optimizing segmentation correction and layout decisions, outputting the optimal reconstruction scheme, and feeding the optimal reconstruction scheme back to the physical workshop for layout adjustment.

[0102] The present invention provides a computer device, including a memory and a processor. The memory stores a program, and when the program is executed by the processor, the processor performs the steps of the above-described method for implementing a discrete workshop reconstruction digital twin.

[0103] According to the disclosed embodiments, the computer device can communicate with one or more external devices (e.g., keyboard, pointing device, Bluetooth communication, etc.) or with any device that enables the computing device to communicate with one or more other computing devices (e.g., router, demodulator, etc.).

[0104] The present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method for implementing a discrete workshop reconstructed digital twin.

[0105] According to the disclosed embodiments, the storage medium can be a non-volatile computer-readable storage medium, such as, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, the storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0106] The above description, in conjunction with specific preferred embodiments, provides a more detailed explanation of the present invention. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for implementing digital twins in discrete workshop reconstruction, characterized in that, Including the following steps: Collect multimodal data in the workshop, including the location of each piece of equipment, the flow path of materials, and the operating status of the equipment, and construct a workshop layout image dataset; Initial instance segmentation is performed on the workshop layout image to extract the original segmentation mask, original category label and visual features of the equipment, and to construct the original segmentation results of each piece of equipment in the workshop. Construct and load an ontology knowledge base that stores equipment attribute knowledge, spatial relationship knowledge, and process constraint knowledge, and obtain the matching degree between the original segmentation result and the ontology knowledge base. If it is lower than the threshold, trigger the correction process, use ontology reasoning rules to correct the original segmentation result, and generate the corrected segmentation result. The corrected segmentation results are semantically mapped to the ontology knowledge base to establish semantic associations for equipment process layout, and candidate layout schemes are proposed based on the semantic information in the semantic associations. In a digital twin environment, candidate layout schemes are simulated in multiple scenarios. The ontology reasoning rules are dynamically updated based on the simulation error to optimize segmentation correction and layout decisions. The optimal reconstruction scheme is then output and fed back to the physical workshop for layout adjustment.

2. The method for implementing a digital twin of discrete workshop reconstruction as described in claim 1, characterized in that, The process involves obtaining the matching degree between the original segmentation result and the ontology knowledge base. If the degree is lower than a threshold, a correction process is triggered. The original segmentation result is corrected using ontology inference rules to generate a corrected segmentation result. Specifically: Based on the target category and location in the original segmentation results, the corresponding equipment attribute knowledge, spatial relationship knowledge and process constraint knowledge are retrieved from the ontology knowledge base through the retrieval function; Based on ontology reasoning rules, the original segmentation results are verified in multiple dimensions. If the results are below the threshold, a correction process is triggered. Ontology reasoning rules are used to correct the category misjudgment, boundary ambiguity, and target omission in the original segmentation results, and a corrected segmentation result is generated.

3. The method for implementing a digital twin of discrete workshop reconstruction as described in claim 2, characterized in that, The original segmentation result is validated in multiple dimensions according to ontology reasoning rules. If the result is below a threshold, a correction process is triggered. The ontology reasoning rules are used to correct for misclassification, blurred boundaries, and missed target detection in the original segmentation result. Specifically: If the target category does not match the device attribute knowledge, it is corrected to a revised target category label based on the device structural characteristics. ; If the original segmentation mask If the corresponding device boundaries do not conform to spatial relationship knowledge, then adjust the original segmentation mask to the corrected segmentation mask. ; If small targets are missed, supplement them. to ,in For the added new target, For the boundary information of the new target, This provides category information for new targets; small targets are those with a pixel area smaller than a threshold.

4. The method for implementing a digital twin of discrete workshop reconstruction as described in claim 1, characterized in that, The process of dynamically updating the ontology inference rules based on simulation errors and optimizing segmentation correction and layout decisions specifically involves: By combining simulation errors, the ontology inference rules are optimized using an ontology rule update function, and then converted into physical entity adjustment instructions through an inverse mapping function to obtain a dynamic optimization scheme for the workshop layout.

5. The method for implementing a digital twin of discrete workshop reconstruction as described in claim 4, characterized in that, The simulation error is specifically: Simulations are performed on the potential layout schemes within a digital twin environment. A performance metric set, including production efficiency (Eff), logistics cost (Cost), and flexibility (Flex), is obtained using a performance index calculation function. P ; like P If the optimization objective is not achieved, a new solution is generated by iteratively updating the function, and the simulation is repeated until the optimal solution is obtained. t The simulation error is obtained by comparing the simulation output at any given time with the measured data from the physical workshop at the same time.

6. The method for implementing a digital twin of discrete workshop reconstruction as described in claim 1, characterized in that, The initial instance segmentation of the workshop layout image involves extracting the original segmentation mask, original category label, and visual features of the equipment to construct the original segmentation results for each piece of equipment in the workshop. Specifically: Visual features of the workshop layout image are extracted using ResNet in the Mask-RCNN model, and visual features at different levels are fused to obtain a fused feature map. On the fused feature map, set the anchor box scale set and aspect ratio set, and perform classification and regression processing to obtain the category score and candidate box coordinates. Then, filter overlapping candidate boxes to obtain the candidate region set. For the candidate region set, feature blocks are extracted using RoI Align through bilinear interpolation, and pixel-level segmentation of workshop equipment is performed using a dual-branch head to obtain the original segmentation results of each piece of equipment in the workshop; the original segmentation results include the original segmentation mask and the original category label.

7. A digital twin implementation system for discrete workshop reconfiguration, characterized in that, include: The data acquisition module is used to collect multimodal data in the workshop, including the location of each piece of equipment, the flow path of materials, and the operating status of the equipment, and to build a workshop layout image dataset. The segmentation module is used to perform initial instance segmentation on the workshop layout image, extract the original segmentation mask, original category label and visual features of the equipment, and construct the original segmentation results of each piece of equipment in the workshop. The correction module is used to build and load an ontology knowledge base that stores equipment attribute knowledge, spatial relationship knowledge, and process constraint knowledge, and obtain the matching degree between the original segmentation result and the ontology knowledge base. If it is lower than the threshold, the correction process is triggered to correct the original segmentation result using ontology reasoning rules and generate the corrected segmentation result. The mapping module is used to semantically map the corrected segmentation results with the ontology knowledge base, establish semantic associations for equipment process layout, and formulate candidate layout schemes based on the semantic information in the semantic associations. The adjustment module is used to perform multi-scenario simulations of candidate layout schemes in a digital twin environment, dynamically update the ontology inference rules based on simulation errors, optimize segmentation correction and layout decisions, output the optimal reconstruction scheme, and feed the optimal reconstruction scheme back to the physical workshop for layout adjustment.

8. A computer device, characterized in that, The system includes a memory and a processor, wherein the memory stores a program that, when executed by the processor, causes the processor to perform the steps of a discrete workshop reconstruction digital twin implementation method as described in any one of claims 1 to 6.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the discrete workshop reconstruction digital twin implementation method according to any one of claims 1 to 6.

Images