An engineering design auxiliary method based on multi-source data fusion

Abstract

This invention discloses an engineering design assistance method based on multi-source data fusion, comprising the following steps: Step 1: Obtaining a fragment set; Step 2: Obtaining a conflict development map; Step 3: Converting the conflict development map into a structured input sequence and inputting it into an improved MaskFormer model, constructing a conflict potential field distribution through a cross-source conflict potential field reflection mechanism, and obtaining a conflict structural unit set; Step 4: Determining the adjudication seed; Step 5: Constructing a manifold topological tension field and performing selective freezing; Step 6: Fusion updating of geometric structural fragments to generate candidate design schemes; Step 7: Matching and verifying the candidate design schemes and performance region fragments to obtain an optimized design scheme; Step 8: Performing consistency screening by combining the optimized design scheme and design intent expression fragments, and outputting the engineering design assistance result. This invention combines the improved MaskFormer model with the manifold topological tension field to achieve efficient engineering design assistance.

Description

Technical Field

[0001] This invention relates to the field of engineering design technology, and in particular to an engineering design assistance method based on multi-source data fusion. Background Technology

[0002] With the continuous improvement of computer-aided design technology and the increasing demand for modeling complex engineering systems, engineering design assistance methods that integrate multi-source heterogeneous data have gradually become an important research direction for improving design efficiency and quality. Existing engineering design processes typically rely on a comprehensive judgment based on multiple data sources, such as CAD drawings, simulation analysis results, and design experience documents. However, the following problems commonly exist in practical applications:

[0003] Data from different sources exhibit significant differences in expression and semantics. CAD drawings emphasize geometric structural information, simulation results highlight performance distribution characteristics, while design texts reflect design intent. Existing methods struggle to map this multi-source information to a unified representation space, making direct comparison and fusion of data difficult and impacting the consistency of overall design decisions. Conflict relationships between multi-source data often rely on human experience for identification and judgment, lacking a quantifiable conflict modeling mechanism. This makes it difficult to accurately locate design boundary changes and promptly identify key conflict areas, increasing design iteration costs. During design optimization, existing methods often make uniform adjustments to the overall structure, lacking constraint and control mechanisms for local key structures. This can lead to unnecessary modifications to already stable structures, reducing design stability. Furthermore, after generating candidate design schemes, existing design-aid technologies typically evaluate based on only a single performance index, lacking a multi-dimensional consistency verification mechanism that combines performance results with design intent. This makes it difficult to ensure that the final design meets both performance requirements and original design goals, limiting the practicality and intelligence level of engineering design aid systems.

[0004] Therefore, how to provide an engineering design assistance method based on multi-source data fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an engineering design assistance method based on multi-source data fusion. This invention combines an improved MaskFormer model with manifold topological tension fields, performs unified modeling and conflict analysis on multi-source data, accurately extracts conflicting structural units, and achieves selective freezing and local adjustment of fragments based on tension distribution. This optimizes the design while ensuring structural stability, and also integrates performance matching and design intent. Figure 1 Consistency verification significantly improves the accuracy, reliability, and overall efficiency of design results.

[0006] An engineering design assistance method based on multi-source data fusion according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Obtain CAD drawing data, simulation result data, and design text data. Extract geometric structure fragments from the CAD drawing data, performance region fragments from the simulation result data, and design intent expression fragments from the design text data. Add source identifiers and operating condition labels to each fragment to obtain a fragment set.

[0008] Step 2: Construct conflict association relationships between fragments based on the fragment set to obtain a conflict rendering map;

[0009] Step 3: Convert the conflict development map into a structured input sequence and input it into the improved MaskFormer model. Construct the conflict potential field distribution through the cross-source conflict potential field reflection mechanism to obtain the conflict structure unit set;

[0010] Step 4: Based on the set of conflicting structural elements, filter out conflicting structural elements that alter the design boundary and determine the decision seed;

[0011] Step 5: Extract the associated fragments that are close to the adjudication seed from the fragment set, construct the manifold topological tension field and perform selective freezing to obtain frozen fragments and unfrozen fragments;

[0012] Step 6: Based on frozen and unfrozen fragments, merge and update the geometric fragments to generate candidate design schemes;

[0013] Step 7: Combine the candidate design schemes with performance region fragments for matching and verification to obtain the optimized design scheme;

[0014] Step 8: Combine the optimized design scheme and design intent expression fragments for consistency screening, and output engineering design auxiliary results.

[0015] Optionally, step one specifically includes:

[0016] Acquire CAD drawing data and parse the CAD drawing data to divide the geometric entity into multiple independent structural units. Extract the shape information, size information and connection relationship of each independent structural unit as a geometric structural fragment.

[0017] Acquire simulation result data, divide the simulation result data into regions, extract the corresponding result information and spatial location of each local result region, and use them as performance region fragments;

[0018] Acquire design text data and perform sentence parsing on the design text data. Break down the sentences including design decisions into single semantic expression units, extract the corresponding design intent information from the single semantic expression units, and use them as design intent expression fragments.

[0019] The geometric structure fragments, performance region fragments, and design intent expression fragments are respectively affixed with source identifiers and operating condition tags. The source identifiers are used to identify the data source to which the fragment belongs, and the operating condition tags are used to identify the operating conditions corresponding to the fragment.

[0020] By summarizing the various fragments with added source identifiers and operating condition labels, a fragment set is obtained.

[0021] Optionally, step two specifically includes:

[0022] The fragment set is grouped according to the working condition label, and association analysis is performed only between fragments with the same working condition label.

[0023] Extract shape, size, and connection information from geometric structure fragments to construct structural feature vectors; extract result information and spatial location information from performance region fragments to construct performance feature vectors; extract keywords and word order information from design intent expression fragments to construct semantic feature vectors.

[0024] The structural feature vector, performance feature vector, and semantic feature vector are input into the multilayer perceptron model, and linear transformation and ReLU activation function nonlinear transformation are performed sequentially. After layer-by-layer processing, standard feature vectors with consistent dimensions are obtained.

[0025] Within the same group, the difference is calculated for the standard feature vectors of any two fragments. The values ​​of the corresponding dimensions of the two standard feature vectors are subtracted to obtain the difference vector. The absolute values ​​of the differences in each dimension of the difference vector are summed, and the summation result is divided by the number of vector dimensions to obtain the average difference value.

[0026] The mean difference between all pairs of fragments in a group is summed and divided by the total number of mean differences to obtain the within-group baseline difference.

[0027] Subtract each mean difference value from the within-group baseline difference value to obtain the deviation value of each mean difference value;

[0028] The variance is obtained by squaring each deviation value and summing the results, then dividing the sum by the total number of average deviation values. The variance is then squared to obtain the within-group variance fluctuation value.

[0029] The intra-group baseline difference value is added to the intra-group difference fluctuation value to obtain the conflict determination threshold;

[0030] The average difference between any two fragments is compared with the conflict determination threshold. When the average difference is greater than the conflict determination threshold, it is determined that there is a conflict relationship between the two fragments.

[0031] By treating fragments as nodes, connecting conflict relationships as undirected edges, and using the average difference value as edge weights, a conflict rendering map is obtained.

[0032] Optionally, converting the conflict imaging map into a structured input sequence specifically involves:

[0033] Extract the standard feature vector corresponding to each node in the conflict development map, and then convert the source identifier and operating condition label into numerical codes and append them to the standard feature vector to generate the node input vector;

[0034] The node input vector is input into a fully connected layer for feature mapping, and then subjected to a non-linear transformation using the GELU activation function to obtain the node embedding vector;

[0035] Construct adjacency relationships based on the undirected edge connections between nodes, and record the node numbers and corresponding edge weights connected to each node.

[0036] The node embedding vector, adjacent node numbers, and corresponding edge weights are concatenated in order to obtain the node structured vector.

[0037] By concatenating the structured vectors of all nodes, a structured input sequence is obtained.

[0038] Optionally, the improved MaskFormer model specifically includes a feature encoding layer, a potential field calculation layer, a query generation layer, and a mask prediction layer;

[0039] The feature coding layer receives a structured input sequence and performs three sets of linear transformations on the structured vector of each node to obtain the corresponding query vector, key vector, and value vector, respectively.

[0040] Multiply the query vector corresponding to any node with the key vectors corresponding to all nodes dimension by dimension and then add them together to obtain the correlation value between the current node and each node.

[0041] For each correlation value, a power operation with the natural constant e as the base is performed to obtain the corresponding correlation index;

[0042] Divide each correlation index by the sum of all correlation indices to obtain the corresponding normalized weight value.

[0043] Each normalized weight value is multiplied element-wise with the value vector of the corresponding node, and all multiplication results are summed element-wise to obtain the node feature representation of the corresponding node. A set of node feature representations is constructed based on all node feature representations.

[0044] The potential field calculation layer receives the set of node feature representations, introduces a cross-source conflict potential field reflection mechanism to fuse the node feature representations and edge weights, and adds up all the edge weights corresponding to each node to obtain the sum of the node edge weights.

[0045] The sum of the values ​​of each dimension in the feature representation of each node is obtained by adding the node feature sum, and the sum of the node edge weights is multiplied by the sum of the node features to obtain the node conflict intensity value.

[0046] Divide the conflict intensity value of each node by the sum of the conflict intensity values ​​of all nodes to obtain the corresponding normalized conflict intensity value. Based on the normalized conflict intensity values ​​of all nodes, a conflict potential field distribution is constructed.

[0047] The query generation layer receives the conflict potential field distribution and the node feature representation set, multiplies each dimension of the query vector by the normalized conflict intensity value of the corresponding node in the conflict potential field distribution to obtain the weighted query component of the corresponding node, and adds all the weighted query components element by element to obtain the weighted query vector.

[0048] The weighted query vector is added element-wise to each node feature representation in the node feature representation set to obtain the updated query vector set.

[0049] The mask prediction layer receives the updated query vector set and the node feature representation set. For each updated query vector, it performs a dimension-wise multiplication and summation with each node feature representation in the node feature representation set to obtain the node matching value of the corresponding node.

[0050] For each node matching value, perform an exponentiation operation with the natural constant e as the base to obtain the corresponding matching index value, and divide each matching index value by the sum of all matching indices to obtain the corresponding normalized response value.

[0051] The average response value is obtained by summing the normalized response values ​​and dividing by the number of response values. Nodes with normalized response values ​​greater than the average response value are identified as conflict structural elements, thus obtaining a set of conflict structural elements.

[0052] Optionally, step four specifically involves:

[0053] For each conflict structural element in the conflict structural element set, extract the corresponding normalized conflict intensity value and normalized response value;

[0054] Add the normalized conflict intensity value to the normalized response value to obtain the boundary influence value of the corresponding conflict structural unit;

[0055] The average boundary influence value is obtained by summing all the boundary influence values ​​in the set of conflicting structural elements and dividing by the number of conflicting structural elements.

[0056] Conflicting structural elements whose boundary influence value is greater than the average boundary influence value are identified as conflicting structural elements that alter the design boundary.

[0057] The conflicting structural element with the largest boundary influence value among the conflicting structural elements that change the design boundary is identified as the adjudication seed.

[0058] Optionally, step five specifically includes:

[0059] Using the fragment corresponding to the adjudication seed as the reference fragment, for each fragment in the fragment set, calculate the difference vector between the node feature representation corresponding to the current fragment and the node feature representation corresponding to the reference fragment, and add the values ​​of each dimension in the difference vector to obtain the associated difference value.

[0060] Add all the association difference values ​​together and divide by the number of fragments in the fragment set to get the average association difference value. Determine the fragments whose association difference value is less than or equal to the average association difference value as the association fragments that are close to the adjudication seed, and obtain the association fragment set.

[0061] For each associated fragment in the associated fragment set, extract the corresponding node feature representation, and add the values ​​of each dimension in the node feature representation to obtain the fragment feature value;

[0062] For any two associated fragments in the associated fragment set, calculate the difference between the feature values ​​of the two fragments, and perform an absolute value operation on the difference to obtain the fragment tension value;

[0063] For each associated fragment, the tension values ​​of all fragments corresponding to the associated fragment are added together to obtain the overall tension value of the associated fragment.

[0064] Divide the overall tension value of each associated fragment by the sum of all overall tension values ​​to obtain the corresponding normalized tension value. Based on all normalized tension values, construct the manifold topological tension field.

[0065] The average tension value is obtained by summing all normalized tension values ​​and dividing by the number of associated fragments in the associated fragment set. Associated fragments with normalized tension values ​​less than the average tension value are identified as frozen fragments and frozen. Associated fragments with normalized tension values ​​greater than or equal to the average tension value are identified as unfrozen fragments.

[0066] Optionally, step six specifically includes:

[0067] For each unfrozen fragment, extract the corresponding normalized tension value, and multiply the normalized tension value with the node feature representation corresponding to the unfrozen fragment dimension by dimension to obtain the fragment adjustment vector of the corresponding unfrozen fragment.

[0068] The element-wise summation of all fragment adjustment vectors yields the fused adjustment vector;

[0069] Determine the unfrozen geometry fragments and frozen geometry fragments;

[0070] The fusion adjustment vector is added dimension by dimension to the node feature representation of each unfrozen geometric fragment to obtain the updated node feature representation of the unfrozen geometric fragment.

[0071] Preserve the node feature representation of the frozen geometry fragments unchanged;

[0072] The updated node feature representations of non-frozen geometric fragments are combined with the node feature representations of frozen geometric fragments to generate candidate design schemes.

[0073] Optionally, step seven specifically includes:

[0074] Extract the node feature representations corresponding to all performance region fragments, add them element by element and divide by the number of performance region fragments to obtain a unified performance feature representation;

[0075] The feature representations of each node in the candidate design scheme are subtracted dimension by dimension from the unified performance feature representation to obtain the corresponding feature difference vector;

[0076] The absolute values ​​of each dimension in each feature difference vector are calculated and then summed to obtain the corresponding matching deviation value. The average matching deviation value is then calculated.

[0077] The node features whose matching deviation value is less than or equal to the average matching deviation value in the candidate design scheme are retained to obtain the optimized design scheme.

[0078] Optionally, step eight specifically includes:

[0079] Extract the node feature representations corresponding to the design intent expression fragments, add them element by element and divide by the number of design intent expression fragments to obtain a unified intent feature representation;

[0080] The node feature representation corresponding to the optimized design scheme is multiplied dimension by dimension with the unified intent feature representation to obtain the product value of the corresponding dimension;

[0081] The number of dimensions with product values ​​greater than or equal to zero is counted, and then divided by the number of dimensions represented by node features to obtain the consistency ratio value. The average consistency ratio value is then calculated.

[0082] The node features in the optimized design scheme whose consistency ratio is greater than or equal to the average consistency ratio are retained and identified as auxiliary results for engineering design.

[0083] The beneficial effects of this invention are:

[0084] This invention combines an improved MaskFormer model with a manifold topological tension field. It constructs a conflict potential field distribution through a cross-source conflict potential field reflection mechanism and further calculates normalized tension values, achieving efficient extraction of conflict structural units and selective freezing of fragments. This allows for constraint control of the design space and precise positioning of locally adjustable regions. Furthermore, it unifies the parsing and fragmentation of CAD drawing data, simulation result data, and design text data, and combines source identifiers and operating condition labels to achieve structured representation of multi-source heterogeneous data. In the conflict identification stage, a conflict rendering map is constructed based on standard feature vectors, and the improved MaskFormer model is used to extract conflict structural units. In the key conflict screening stage, boundary influence values ​​are constructed using normalized conflict intensity values ​​and normalized response values, and the average boundary influence value is used to screen out key conflicts. The design process involves identifying conflicting structural elements that alter the design boundary and determining adjudication seeds. During the design generation phase, normalized tension values ​​drive the generation of fragment adjustment vectors for unfrozen fragments. These adjustment vectors are then fused to obtain a fused adjustment vector, which is applied to the node feature representations of unfrozen geometric fragments for fusion updates, while keeping frozen geometric fragments unchanged, thus generating candidate design schemes. In the result verification phase, a unified performance feature representation is used to calculate and filter the matching deviation values ​​of the candidate design schemes to obtain optimized schemes. Furthermore, a unified intent feature representation is used to determine the consistency ratio between the optimized scheme and the optimized scheme, resulting in the output of engineering design assistance results. This approach, while ensuring structural stability, enables precise design adjustments driven by multi-source information, improving the accuracy, stability, and intelligence of engineering design assistance. Attached Figure Description

[0085] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0086] Figure 1 This is an overall flowchart of an engineering design auxiliary method based on multi-source data fusion proposed in this invention;

[0087] Figure 2 This is a schematic diagram of the improved MaskFormer model structure of an engineering design auxiliary method based on multi-source data fusion proposed in this invention.

[0088] Figure 3This is a flowchart illustrating the construction and selective freezing of a manifold topological tension field based on a multi-source data fusion-based engineering design auxiliary method proposed in this invention. Detailed Implementation

[0089] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0090] refer to Figures 1-3 An engineering design assistance method based on multi-source data fusion includes the following steps:

[0091] Step 1: Obtain CAD drawing data, simulation result data, and design text data. Extract geometric structure fragments from the CAD drawing data, performance region fragments from the simulation result data, and design intent expression fragments from the design text data. Add source identifiers and operating condition labels to each fragment to obtain a fragment set.

[0092] Step 2: Construct conflict relationships between fragments based on the fragment set to obtain a conflict rendering map;

[0093] Step 3: Convert the conflict development map into a structured input sequence and input it into the improved MaskFormer model. Construct the conflict potential field distribution through the cross-source conflict potential field reflection mechanism to obtain the conflict structure unit set;

[0094] Step 4: Based on the set of conflicting structural elements, filter out conflicting structural elements that alter the design boundary and determine the decision seed;

[0095] Step 5: Extract related fragments that are close to the adjudication seed from the fragment set, construct the manifold topological tension field and perform selective freezing to obtain frozen fragments and unfrozen fragments;

[0096] Step 6: Based on frozen and unfrozen fragments, merge and update the geometric fragments to generate candidate design schemes;

[0097] Step 7: Match and verify the candidate design schemes with the performance region fragments to obtain the optimized design scheme;

[0098] Step 8: Combine the optimized design scheme and design intent expression fragments for consistency screening, and output engineering design auxiliary results.

[0099] In this embodiment, step one specifically includes:

[0100] Acquire CAD drawing data and parse the CAD drawing data. Divide the geometric entity into multiple independent structural units. Extract the shape information, size information and connection relationship of each independent structural unit as a geometric structural fragment.

[0101] Acquire simulation result data, divide the simulation result data into regions, extract the corresponding result information and spatial location of each local result region, and use them as performance region fragments;

[0102] Acquire design text data and perform sentence parsing on the design text data. Break down the sentences containing design decisions into single semantic expression units, extract the corresponding design intent information from the single semantic expression units, and use them as design intent expression fragments.

[0103] Add source identifiers and operating condition labels to geometric structure fragments, performance region fragments and design intent expression fragments respectively. The source identifier is used to identify the data source to which the fragment belongs, and the operating condition label is used to identify the operating condition corresponding to the fragment.

[0104] The various fragments with added source identifiers and operating condition labels are aggregated to obtain a fragment set;

[0105] In the specific implementation process, CAD drawing data preferably uses STEP or IGES format files. By parsing the geometric boundary representation information, the overall model is divided into the smallest closed geometric units. The dimensional information extraction accuracy of each independent structural unit is set to 0.01mm, and the connection relationship is encoded through a topological adjacency matrix. Simulation result data uses finite element analysis output files. The simulation result data is divided into regions according to a spatial mesh, with the mesh size set to 2mm×2mm×2mm. The corresponding result value range is extracted for each mesh region, and the result information is represented in the form of maximum and average values. Design text data is parsed using a dependency parsing-based method, breaking down each design statement into... The data is divided into single semantic expression units with a length of no more than 20 words, and action words and target objects are labeled to extract design intent information. The source identifier is represented by a fixed encoding method, where CAD drawing data is encoded as 1, simulation result data is encoded as 2, and design text data is encoded as 3. The working condition label is represented in the form of a multi-dimensional vector, including load value, temperature value, and boundary constraint type. The load value is set to a range of 0 to 10000N, the temperature range is set to -40℃ to 200℃, and the boundary constraint type is represented by an integer encoding. After the above processing is completed, all fragments are uniformly numbered according to the source identifier and working condition label, and sorted in chronological order to form a fragment set with a unified structure.

[0106] In this embodiment, step two specifically includes:

[0107] The fragment set is grouped according to the working condition label, and association analysis is performed only between fragments with the same working condition label;

[0108] Extract shape, size, and connection information from geometric structure fragments to construct structural feature vectors; extract result information and spatial location information from performance region fragments to construct performance feature vectors; extract keywords and word order information from design intent expression fragments to construct semantic feature vectors.

[0109] Specifically, the structural feature vectors of geometric fragments are constructed using an encoding method based on the PointNet network. The point cloud data corresponding to each geometric fragment is used as input, and the PointNet network extracts features from the point set, outputting a fixed-dimensional structural feature vector. The performance feature vectors of performance region fragments are constructed using an encoding method based on a convolutional neural network. The result distribution data corresponding to the performance region is mapped to a regular grid as input, and the spatial distribution features are extracted using a convolutional neural network, outputting a fixed-dimensional performance feature vector. The semantic feature vectors of design intent expression fragments are constructed using an encoding method based on the BERT model. The design statement is input into a pre-trained BERT model, and its encoding layer outputs a semantic feature vector. Through these three encoding methods, different types of fragments directly obtain their corresponding feature vectors after processing by their respective network models.

[0110] The structural feature vector, performance feature vector, and semantic feature vector are input into the multilayer perceptron model, and linear transformation and nonlinear transformation with ReLU activation function are performed sequentially. After layer-by-layer processing, standard feature vectors with consistent dimensions are obtained.

[0111] Within the same group, the difference is calculated for the standard feature vectors of any two fragments. The values ​​of the corresponding dimensions of the two standard feature vectors are subtracted to obtain the difference vector. The absolute values ​​of the differences in each dimension of the difference vector are summed, and the sum is divided by the number of vector dimensions to obtain the average difference value.

[0112] The mean difference between all pairs of fragments in a group is summed and divided by the total number of mean differences to obtain the within-group baseline difference.

[0113] Subtract each mean difference value from the within-group baseline difference value to obtain the deviation value of each mean difference value;

[0114] Square each deviation value and sum them up. Divide the sum by the total number of average deviation values ​​to obtain the variance value. Take the square root of the variance value to obtain the within-group variance fluctuation value.

[0115] The threshold for conflict determination is obtained by adding the intra-group baseline difference value to the intra-group difference fluctuation value.

[0116] The average difference between any two fragments is compared with the conflict determination threshold. When the average difference is greater than the conflict determination threshold, it is determined that there is a conflict relationship between the two fragments.

[0117] By treating fragments as nodes, connecting conflict relationships as undirected edges, and using the average difference value as the edge weight, a conflict rendering map is obtained.

[0118] In this invention, by constructing feature vectors from fragments from different sources and mapping them to a standard feature space of consistent dimensions, geometric structure information, performance result information, and design semantic information can be compared and analyzed within the same mathematical space, thereby avoiding the incomparability problem caused by differences in expression forms between multi-source data. Furthermore, by calculating the difference in each dimension of the standard feature vectors and obtaining the average difference value, a quantitative representation of the degree of difference between fragments is achieved, transforming the conflict relationship from a qualitative judgment into a calculable numerical judgment, thus improving the objectivity and stability of conflict identification. By introducing the intra-group benchmark difference value and the intra-group difference fluctuation value to construct a conflict judgment threshold, the threshold can be automatically adjusted according to the data distribution within the current group, avoiding the insufficient adaptability problem caused by a fixed threshold, thereby improving robustness under different working conditions. At the same time, the conflict relationship is expressed in the form of an undirected weighted graph, with the average difference value as the edge weight, so that the conflict development map can intuitively reflect the conflict intensity distribution between fragments, providing a reliable data foundation for design decisions based on the conflict structure, thereby improving the accuracy and efficiency of engineering design assistance.

[0119] In this embodiment, the conflict imaging map is converted into a structured input sequence, specifically as follows:

[0120] Extract the standard feature vector corresponding to each node in the conflict development map, and then convert the source identifier and operating condition label into numerical codes and append them to the standard feature vector to generate the node input vector;

[0121] The node input vector is input into a fully connected layer for feature mapping, and then subjected to a non-linear transformation using the GELU activation function to obtain the node embedding vector;

[0122] Construct adjacency relationships based on the undirected edge connections between nodes, and record the node numbers and corresponding edge weights connected to each node.

[0123] The node embedding vector, adjacent node numbers, and corresponding edge weights are concatenated in order to obtain the node structured vector.

[0124] Concatenate all the structured vectors of the nodes to obtain the structured input sequence;

[0125] In this invention, by converting conflict-revealing maps into structured input sequences, a unified representation of graph-structured data to serialized data is achieved, enabling the improved MaskFormer model to directly process multi-source conflict information. Node input vectors are mapped through fully connected layers and subjected to nonlinear transformation using the GELU activation function, effectively enhancing feature representation capabilities and giving fragments from different sources stronger discriminativeness and separability in the same feature space. By introducing adjacency relationships and edge weights, conflict association information between fragments is embedded into node representations, allowing structural information and conflict intensity information to participate in modeling simultaneously. The concatenation and connection of node structured vectors enables the original graph structure to be transformed into a serial structure without relying on a complex graph computation framework, thereby reducing model input complexity and improving processing efficiency. The overall scheme achieves efficient expression and compact encoding of multi-source conflict relationships while ensuring information integrity, providing a stable input foundation for conflict structural unit extraction.

[0126] In this embodiment, the improved MaskFormer model specifically includes a feature encoding layer, a potential field calculation layer, a query generation layer, and a mask prediction layer.

[0127] The feature coding layer receives a structured input sequence and performs three sets of linear transformations on the structured vector of each node to obtain the corresponding query vector, key vector, and value vector, respectively.

[0128] Multiply the query vector corresponding to any node with the key vectors corresponding to all nodes dimension by dimension and then add them together to obtain the correlation value between the current node and each node.

[0129] For each correlation value, a power operation with the natural constant e as the base is performed to obtain the corresponding correlation index;

[0130] Divide each correlation index by the sum of all correlation indices to obtain the corresponding normalized weight value.

[0131] Each normalized weight value is multiplied element-wise with the value vector of the corresponding node, and all multiplication results are summed element-wise to obtain the node feature representation of the corresponding node. A set of node feature representations is constructed based on all node feature representations.

[0132] The potential field calculation layer receives the set of node feature representations and introduces a cross-source conflict potential field reflection mechanism to fuse the node feature representations and edge weights. The total edge weights of each node are added together to obtain the total node edge weights.

[0133] The sum of the values ​​of each dimension in the feature representation of each node is obtained by adding the node feature sum, and the sum of the node edge weights is multiplied by the sum of the node features to obtain the node conflict intensity value.

[0134] Divide the conflict intensity value of each node by the sum of the conflict intensity values ​​of all nodes to obtain the corresponding normalized conflict intensity value. Based on the normalized conflict intensity values ​​of all nodes, a conflict potential field distribution is constructed.

[0135] The query generation layer receives the conflict potential field distribution and the set of node feature representations. It multiplies each dimension of the query vector by the normalized conflict intensity value of the corresponding node in the conflict potential field distribution to obtain the weighted query component of the corresponding node. All weighted query components are summed element by element to obtain the weighted query vector.

[0136] The weighted query vector is added element-wise to each node feature representation in the node feature representation set to obtain the updated query vector set.

[0137] The mask prediction layer receives the updated set of query vectors and the set of node feature representations. For each updated query vector, it performs a dimension-wise multiplication and summation with each node feature representation in the set of node feature representations to obtain the node matching value of the corresponding node.

[0138] For each node matching value, perform an exponentiation operation with the natural constant e as the base to obtain the corresponding matching index value, and divide each matching index value by the sum of all matching indices to obtain the corresponding normalized response value.

[0139] The average response value is obtained by summing the normalized response values ​​and dividing by the number of response values. Nodes with normalized response values ​​greater than the average response value are identified as conflict structural elements, thus obtaining a set of conflict structural elements.

[0140] In this invention, the improved MaskFormer model inherits the basic framework of existing MaskFormer models based on encoding, querying, and mask prediction. It performs global interactive modeling of input features through the encoding layer and extracts the target structure through a query mechanism, thus maintaining MaskFormer's advantages in unified modeling and end-to-end processing. Furthermore, the model's input format and internal computation path are improved for adaptability to multi-source engineering data. The traditional input for regular grid data is replaced with a structured input sequence composed of node structured vectors, enabling the model to directly process graph-structured data containing topological relationships and conflict intensity information. In the feature encoding layer, query vectors, key vectors, and value vectors are explicitly constructed, and correlation calculation is performed using dimension-wise multiplication and exponential normalization, making the association calculation process between nodes more efficient. To enhance model stability under non-uniform data distribution, a clear numerical expression is provided. A cross-source conflict potential field reflection mechanism is introduced in the potential field calculation layer, fusing edge weights and node features. The node conflict intensity value is obtained by multiplying the sum of node edge weights with the sum of node features, and then normalized to form a conflict potential field distribution. This achieves intensity mapping of multi-source conflict information in a unified space, enabling the model to highlight key nodes with high conflict levels. In the query generation layer, the conflict potential field distribution is introduced into the query vector update process, directly modulating the query vector with conflict intensity during generation, thereby improving the query vector's responsiveness to high-conflict regions. In the mask prediction layer, the matching results of the updated query vector and node features are exponentially normalized, and conflict structural units are generated using a partitioning method based on average response values, making the structural partitioning process adaptive. Through these improvements, the MaskFormer model can efficiently focus and structurally extract conflict information under multi-source heterogeneous data conditions, improving the accuracy and stability of key conflict identification in engineering design assistance.

[0141] In this embodiment, step four specifically includes:

[0142] For each conflict structural element in the conflict structural element set, extract the corresponding normalized conflict intensity value and normalized response value;

[0143] Add the normalized conflict intensity value to the normalized response value to obtain the boundary influence value of the corresponding conflict structural unit;

[0144] The average boundary influence value is obtained by summing all the boundary influence values ​​in the set of conflicting structural elements and dividing by the number of conflicting structural elements.

[0145] Conflicting structural elements whose boundary influence value is greater than the average boundary influence value are identified as conflicting structural elements that alter the design boundary.

[0146] The conflicting structural element with the largest boundary influence value among the conflicting structural elements that change the design boundary is identified as the adjudication seed.

[0147] This invention integrates the normalized conflict intensity value and normalized response value corresponding to each conflicting structural unit, mapping multi-source conflict information into a unified boundary influence value, thus achieving a comprehensive quantitative expression of conflict intensity and structural response. The normalized conflict intensity value reflects the relative proportion of conflict in the whole, while the normalized response value reflects the node response degree under the query mechanism. The sum of the two makes the boundary influence value possess both global and local response characteristics. By averaging all boundary influence values, an adaptive judgment benchmark is constructed, enabling conflicting structural units of different sizes or distributions to be screened under a unified standard. Further comparison and screening based on the average boundary influence value effectively eliminates conflicting structural units with weaker influence, thereby highlighting key units that have a significant impact on the design boundary. Finally, by selecting the conflicting structural unit with the largest boundary influence value as the adjudication seed, the most influential conflict location is accurately located, providing a clear basis for engineering design adjustments and improving the pertinence and efficiency of design assistance.

[0148] In this embodiment, step five specifically includes:

[0149] Using the fragment corresponding to the adjudication seed as the reference fragment, for each fragment in the fragment set, calculate the difference vector between the node feature representation corresponding to the current fragment and the node feature representation corresponding to the reference fragment, and add the values ​​of each dimension in the difference vector to obtain the associated difference value.

[0150] Add all the association difference values ​​together and divide by the number of fragments in the fragment set to get the average association difference value. Determine the fragments whose association difference value is less than or equal to the average association difference value as the association fragments that are close to the adjudication seed, and obtain the association fragment set.

[0151] For each associated fragment in the associated fragment set, extract the corresponding node feature representation, and add the values ​​of each dimension in the node feature representation to obtain the fragment feature value;

[0152] For any two associated fragments in the associated fragment set, calculate the difference between the feature values ​​of the two fragments, and perform an absolute value operation on the difference to obtain the fragment tension value;

[0153] For each associated fragment, the tension values ​​of all fragments corresponding to the associated fragment are added together to obtain the overall tension value of the associated fragment.

[0154] Divide the overall tension value of each associated fragment by the sum of all overall tension values ​​to obtain the corresponding normalized tension value. Based on all normalized tension values, construct the manifold topological tension field.

[0155] Add all normalized tension values ​​and divide by the number of associated fragments in the associated fragment set to obtain the average tension value. Associated fragments with normalized tension values ​​less than the average tension value are identified as frozen fragments and frozen. Associated fragments with normalized tension values ​​greater than or equal to the average tension value are identified as unfrozen fragments.

[0156] In this invention, the freezing operation is used to constrain and control some fragments. The node feature representations that are determined to be frozen fragments are marked, and the corresponding fragments are masked in subsequent steps, that is, they do not participate in the mapping fusion calculation and update. Correspondingly, the non-frozen fragments participate in the subsequent mapping fusion and update as adjustable parts, which are used to adjust and optimize the design scheme.

[0157] By constructing a manifold topological tension field, the differences between related fragments are transformed into a tension distribution, so that each fragment corresponds to a normalized tension value. The normalized tension value reflects the sensitivity of the fragment to changes in the overall structure. Fragments with smaller normalized tension values ​​correspond to relatively stable structures and are suitable for retention as constraints, while fragments with larger normalized tension values ​​correspond to larger change space and are more suitable for adjustment. Selective freezing based on the manifold topological tension field can make targeted adjustments to local areas while maintaining the stability of the overall structure, thereby avoiding irrelevant fragments from participating in design changes and reducing unnecessary structural disturbances. This method achieves effective constraint and zonal control of the design space, making the design adjustment process directional and controllable, and improving the stability and efficiency of engineering design assistance.

[0158] In this embodiment, step six specifically includes:

[0159] For each unfrozen fragment, extract the corresponding normalized tension value, and multiply the normalized tension value with the node feature representation corresponding to the unfrozen fragment dimension by dimension to obtain the fragment adjustment vector of the corresponding unfrozen fragment.

[0160] The element-wise summation of all fragment adjustment vectors yields the fused adjustment vector;

[0161] Determine the unfrozen geometry fragments and frozen geometry fragments;

[0162] The fusion adjustment vector is added dimension by dimension to the node feature representation of each unfrozen geometric fragment to obtain the updated node feature representation of the unfrozen geometric fragment.

[0163] Preserve the node feature representation of the frozen geometry fragments unchanged;

[0164] The updated node feature representations of the unfrozen geometric fragments are combined with the node feature representations of the frozen geometric fragments to generate candidate design schemes.

[0165] In this invention, the unfrozen fragments are adjusted by normalizing the tension value, enabling each fragment to participate in the design adjustment according to its importance in the manifold topological tension field during the fusion process, thus achieving weighted integration of multi-source information. By fusing the adjustment vectors of each fragment to form a unified fusion adjustment vector, the problem of local inconsistency caused by mutual interference between different fragments is avoided, improving the coordination of the overall design adjustment. During the update process, only the node feature representations of the unfrozen geometric structure fragments are adjusted one by one, while the node feature representations of the frozen geometric structure fragments remain unchanged, so that key structural regions are stably preserved and unnecessary structural fluctuations are reduced. At the same time, by applying the unified fusion adjustment vector to multiple unfrozen geometric structure fragments, the design adjustment has global consistency and directionality, improving the overall rationality and feasibility of candidate design schemes, thereby improving the efficiency and stability of engineering design assistance.

[0166] In this embodiment, step seven specifically includes:

[0167] Extract the node feature representations corresponding to all performance region fragments, add them element by element and divide by the number of performance region fragments to obtain a unified performance feature representation;

[0168] The feature representations of each node in the candidate design scheme are subtracted from the unified performance feature representation dimension by dimension to obtain the corresponding feature difference vector;

[0169] The absolute values ​​of each dimension in each feature difference vector are calculated and then summed to obtain the corresponding matching deviation value. The average matching deviation value is then calculated.

[0170] The node features whose matching deviation value is less than or equal to the average matching deviation value in the candidate design scheme are retained to obtain the optimized design scheme.

[0171] In this invention, a unified performance feature representation is obtained by uniformly fusing the node feature representations corresponding to fragments in multiple performance regions. This enables the performance information of different regions to form a consistent representation within the same feature space, thus providing a unified benchmark for subsequent matching. By calculating the difference between the node feature representations in the candidate design schemes and the unified performance feature representation dimension by dimension, and accumulating the absolute values ​​of the differences, a matching deviation value is obtained, which quantifies the degree of difference between structural features and performance features. Furthermore, by calculating the average matching deviation value as an adaptive judgment benchmark, the node feature representations in the candidate design schemes are screened, retaining only the parts with high consistency with performance features, thereby eliminating structural information with large deviations. This method can improve the matching degree between the design results and performance requirements while ensuring the integrity of the overall structure, so that the final optimized design scheme has better balance and stability in terms of structural rationality and performance consistency.

[0172] In this embodiment, step eight specifically includes:

[0173] Extract the node feature representations corresponding to the design intent expression fragments, add them element by element and divide by the number of design intent expression fragments to obtain a unified intent feature representation;

[0174] The node feature representation corresponding to the optimized design scheme is multiplied with the unified intent feature representation dimension by dimension to obtain the product value of the corresponding dimension;

[0175] The number of dimensions with product values ​​greater than or equal to zero is counted, and then divided by the number of dimensions represented by node features to obtain the consistency ratio value. The average consistency ratio value is then calculated.

[0176] Node features whose consistency ratio is greater than or equal to the average consistency ratio in the optimized design scheme are retained and identified as auxiliary results for engineering design.

[0177] In this invention, a unified intent feature representation is obtained by adding the node feature representations corresponding to multiple design intent expression fragments element by element and dividing by the number. This allows the dispersed design intent information to form a consistent expression in the same feature space, thereby avoiding the deviation caused by a single design statement. On this basis, the node feature representation corresponding to the optimized design scheme is multiplied with the unified intent feature representation in each dimension. The positive or negative relationship of the product value is used to characterize the directional consistency between the design scheme and the design intent in each dimension, thereby achieving an effective evaluation of whether the design result meets the original design goal. Furthermore, the consistency ratio value and the average consistency ratio value are used for screening, retaining only the node feature representations that are consistent with the design intent in most dimensions, making the final result more in line with the design expectation.

[0178] The engineering design assistance results obtained based on the above processing can be directly applied to the actual engineering design process. The filtered node feature representations are mapped back to the corresponding geometric structural fragments to guide structural dimension optimization or local morphological modification and updates. At the same time, the structural update data is written into the computer-aided design software through an interface program, which modifies the geometric entity parameters in the CAD model accordingly, realizing automatic updates of the CAD model. In specific implementation, the update parameters can be written into the parameter fields of the corresponding geometric entities one by one by calling the application programming interface provided by the CAD software, and triggering the model reconstruction operation, so that the updated design model is generated in real time in the CAD environment. At the same time, the update process is recorded as an operation sequence for visualization in the design interface or for confirmation by engineers, thereby realizing the effective feedback and application of engineering design assistance results to the actual CAD model.

[0179] Example 1: To verify the feasibility of the present invention in practice, the present invention was applied to the engineering design optimization process of a complex mechanical structure. The structure contains multiple load-bearing connection areas and stress transition areas. In the initial design stage, simulation results showed that stress concentration existed in some areas, and redundant design existed in some structural areas, resulting in high overall structural weight and unstable local performance. Traditional design methods require multiple rounds of structural modification and simulation verification, which is inefficient and results in large fluctuations.

[0180] In practical applications, the system first acquires CAD drawing data, simulation result data, and design text data. In one processing step, 128 geometric structural fragments are parsed from the CAD drawing data, each corresponding to an independent structural unit; 96 performance region fragments are obtained from the simulation result data; and 54 design intent expression fragments are parsed from the design text data. Source identifiers and operating condition tags are added to all fragments, thus forming a unified fragment set.

[0181] During the construction of the conflict development map, the differences in the standard feature vectors of fragments are calculated. For example, the average difference between two fragments is calculated to be 0.187. Within this working condition group, the baseline difference within the group is 0.094, and the fluctuation value within the group is 0.041. Thus, the conflict determination threshold is 0.135. When the average difference value is greater than this threshold, a conflict relationship is determined to exist. In this implementation process, a total of 462 conflict relationships were identified, and a conflict development map containing 462 undirected edges was constructed.

[0182] In the process of conflict structure extraction, the conflict rendering map is converted into a structured input sequence and input into the improved MaskFormer model. In the feature encoding stage, the correlation between nodes is calculated. For example, after calculating the correlation index between a node and other nodes, the weight value is normalized to approximately 0.032. The node feature representation is obtained by weighted summation. In the potential field calculation process, for a certain node, the sum of its edge weights is 3.74 and the sum of its features is 1.92. The two are multiplied to obtain the node conflict intensity value of 7.18. After further normalization, the normalized conflict intensity value of the node is 0.028. Finally, the conflict potential field distribution is formed, and 17 conflict structure units are extracted.

[0183] During the conflict screening process, the boundary influence value is calculated for each conflict structural unit. For example, the normalized conflict intensity value of a certain unit is 0.036, the normalized response value is 0.041, and the two are added together to obtain the boundary influence value of 0.077. The average boundary influence value among all conflict structural units is 0.052. Six key conflict structural units are screened out, and the unit with the largest boundary influence value (0.089) is determined as the adjudication seed.

[0184] During the construction of the manifold topological tension field, the association difference value is calculated for 128 fragments based on the adjudication seed. For example, if the association difference value of a certain fragment is 2.63, the average association difference value of all fragments is 1.84, and 49 associated fragments are obtained. The fragment tension value is further calculated. For example, if the difference between the eigenvalues ​​of two fragments is 0.65, the cumulative overall tension value of a certain fragment is 8.73. After normalization, the tension value distribution range is 0.011 to 0.067, and the average tension value is 0.032. Finally, 21 frozen fragments and 28 unfrozen fragments are obtained.

[0185] During the design generation process, a normalized tension value is extracted for each unfrozen fragment. For example, if the normalized tension value of a fragment is 0.054, the corresponding node feature representation has a value of 0.82 in one dimension. Multiplying the two together yields an adjustment component of 0.044, forming a fragment adjustment vector. The 28 fragment adjustment vectors are then summed element-wise and divided by the number of unfrozen fragments to obtain a fused adjustment vector, for example, with a value of 0.049 in one dimension. This fused adjustment vector is applied to the unfrozen geometric fragments. For example, if the original node feature representation of a fragment has a value of 0.63 in this dimension, it is updated to 0.679, achieving smooth adjustment. At the same time, the frozen geometric fragments are kept unchanged, thus generating candidate design schemes.

[0186] During the performance matching verification process, the feature representations of the 96 performance region fragment nodes are uniformly processed. For example, if the sum of a certain dimension is 48.6, dividing by 96 yields a uniform performance feature representation of 0.506. The matching deviation value is calculated for each node in the candidate design scheme. For example, if the matching deviation value of a certain node is 1.37, the average matching deviation value among all nodes is 1.02. The feature representations of nodes in the candidate design scheme whose matching deviation value is less than or equal to the average matching deviation value are retained, while the feature representations of nodes whose matching deviation value is greater than the average matching deviation value are removed, thus obtaining the optimized design scheme.

[0187] During the consistency screening process, the 54 design intent expression fragments were unified to obtain intent feature representations; the node feature representations corresponding to the optimized design schemes were multiplied dimension by dimension with the unified intent feature representations to obtain the product values ​​of the corresponding dimensions, and further obtained the consistency ratio values, and the average consistency ratio value was calculated; the node feature representations in the optimized design schemes with consistency ratio values ​​greater than or equal to the average consistency ratio value were retained, and the node feature representations with consistency ratio values ​​less than the average consistency ratio value were removed (i.e., the adjustment amounts that deviate from the design intent were removed), and these were determined as auxiliary results for engineering design.

[0188] In practical applications, the filtered node feature representations are mapped back to geometric structure fragments. For example, if a structural dimension is adjusted from 12.50 to 12.94, or a connection angle is adjusted from 45.0 to 46.2, this is written into the CAD model via an interface program for automatic updates. Comparative results show that under the same design task, the traditional method requires 7 iterations, while this invention requires only 3. The matching deviation value is reduced from approximately 1.95 to 0.98, and the consistency ratio value is increased from 0.65 to 0.88, while the structure remains intact and stable, achieving efficient and controllable engineering design optimization.

[0189] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for engineering design assistance based on multi-source data fusion, characterized in that, Includes the following steps: Step 1: Obtain CAD drawing data, simulation result data, and design text data. Extract geometric structure fragments from the CAD drawing data, performance region fragments from the simulation result data, and design intent expression fragments from the design text data. Add source identifiers and operating condition labels to each fragment to obtain a fragment set. Step 2: Construct conflict association relationships between fragments based on the fragment set to obtain a conflict rendering map; Step 3: Convert the conflict development map into a structured input sequence and input it into the improved MaskFormer model. Construct the conflict potential field distribution through the cross-source conflict potential field reflection mechanism to obtain the conflict structure unit set; Step 4: Based on the set of conflicting structural elements, filter out conflicting structural elements that alter the design boundary and determine the decision seed; Step 5: Extract the associated fragments that are close to the adjudication seed from the fragment set, construct the manifold topological tension field and perform selective freezing to obtain frozen fragments and unfrozen fragments; Step 6: Based on frozen and unfrozen fragments, merge and update the geometric fragments to generate candidate design schemes; Step 7: Combine the candidate design schemes with performance region fragments for matching and verification to obtain the optimized design scheme; Step 8: Combine the optimized design scheme and design intent expression fragments for consistency screening, and output engineering design auxiliary results.

2. The engineering design assistance method based on multi-source data fusion according to claim 1, characterized in that, Step one specifically involves: Acquire CAD drawing data and parse the CAD drawing data to divide the geometric entity into multiple independent structural units. Extract the shape information, size information and connection relationship of each independent structural unit as a geometric structural fragment. Acquire simulation result data, divide the simulation result data into regions, extract the corresponding result information and spatial location of each local result region, and use them as performance region fragments; Acquire design text data and perform sentence parsing on the design text data. Break down the sentences including design decisions into single semantic expression units, extract the corresponding design intent information from the single semantic expression units, and use them as design intent expression fragments. The geometric structure fragments, performance region fragments, and design intent expression fragments are respectively affixed with source identifiers and operating condition tags. The source identifiers are used to identify the data source to which the fragment belongs, and the operating condition tags are used to identify the operating conditions corresponding to the fragment. By summarizing the various fragments with added source identifiers and operating condition labels, a fragment set is obtained.

3. The engineering design assistance method based on multi-source data fusion according to claim 1, characterized in that, Step two specifically involves: The fragment set is grouped according to the working condition label, and association analysis is performed only between fragments with the same working condition label. Extract shape information, size information, and connection relationships from geometric fragments to construct structural feature vectors; Extract performance region fragment information and spatial location information to construct performance feature vectors; Extract keywords and word order information from the sentences expressing the design intent, and construct semantic feature vectors; The structural feature vector, performance feature vector, and semantic feature vector are input into the multilayer perceptron model, and linear transformation and ReLU activation function nonlinear transformation are performed sequentially. After layer-by-layer processing, standard feature vectors with consistent dimensions are obtained. Within the same group, the difference is calculated for the standard feature vectors of any two fragments. The values ​​of the corresponding dimensions of the two standard feature vectors are subtracted to obtain the difference vector. The absolute values ​​of the differences in each dimension of the difference vector are summed, and the summation result is divided by the number of vector dimensions to obtain the average difference value. The mean difference between all pairs of fragments in a group is summed and divided by the total number of mean differences to obtain the within-group baseline difference. Subtract each mean difference value from the within-group baseline difference value to obtain the deviation value of each mean difference value; The variance is obtained by squaring each deviation value and summing the results, then dividing the sum by the total number of average deviation values. The variance is then squared to obtain the within-group variance fluctuation value. The intra-group baseline difference value is added to the intra-group difference fluctuation value to obtain the conflict determination threshold; The average difference between any two fragments is compared with the conflict determination threshold. When the average difference is greater than the conflict determination threshold, it is determined that there is a conflict relationship between the two fragments. By treating fragments as nodes, connecting conflict relationships as undirected edges, and using the average difference value as edge weights, a conflict rendering map is obtained.

4. The engineering design assistance method based on multi-source data fusion according to claim 1, characterized in that, The process of converting the conflict-detection map into a structured input sequence specifically involves: Extract the standard feature vector corresponding to each node in the conflict development map, and then convert the source identifier and operating condition label into numerical codes and append them to the standard feature vector to generate the node input vector; The node input vector is input into the fully connected layer for feature mapping, and then a non-linear transformation is performed through the GELU activation function to obtain the node embedding vector; Construct adjacency relationships based on the undirected edge connections between nodes, and record the node numbers and corresponding edge weights connected to each node. The node embedding vector, adjacent node numbers, and corresponding edge weights are concatenated in order to obtain the node structured vector. By concatenating the structured vectors of all nodes, a structured input sequence is obtained.

5. The engineering design assistance method based on multi-source data fusion according to claim 1, characterized in that, The improved MaskFormer model specifically includes a feature encoding layer, a potential field calculation layer, a query generation layer, and a mask prediction layer. The feature coding layer receives a structured input sequence and performs three sets of linear transformations on the structured vector of each node to obtain the corresponding query vector, key vector, and value vector, respectively. Multiply the query vector corresponding to any node with the key vectors corresponding to all nodes dimension by dimension and then add them together to obtain the correlation value between the current node and each node. For each correlation value, a power operation with the natural constant e as the base is performed to obtain the corresponding correlation index; Divide each correlation index by the sum of all correlation indices to obtain the corresponding normalized weight value. Each normalized weight value is multiplied element-wise with the value vector of the corresponding node, and all multiplication results are summed element-wise to obtain the node feature representation of the corresponding node. A set of node feature representations is constructed based on all node feature representations. The potential field calculation layer receives the set of node feature representations, introduces a cross-source conflict potential field reflection mechanism to fuse the node feature representations and edge weights, and adds up all the edge weights corresponding to each node to obtain the sum of the node edge weights. The sum of the values ​​of each dimension in the feature representation of each node is obtained by adding the node feature sum, and the sum of the node edge weights is multiplied by the sum of the node features to obtain the node conflict intensity value. Divide the conflict intensity value of each node by the sum of the conflict intensity values ​​of all nodes to obtain the corresponding normalized conflict intensity value. Based on the normalized conflict intensity values ​​of all nodes, a conflict potential field distribution is constructed. The query generation layer receives the conflict potential field distribution and the node feature representation set, multiplies each dimension of the query vector by the normalized conflict intensity value of the corresponding node in the conflict potential field distribution to obtain the weighted query component of the corresponding node, and adds all the weighted query components element by element to obtain the weighted query vector. The weighted query vector is added element-wise to each node feature representation in the node feature representation set to obtain the updated query vector set. The mask prediction layer receives the updated query vector set and the node feature representation set. For each updated query vector, it performs a dimension-wise multiplication and summation with each node feature representation in the node feature representation set to obtain the node matching value of the corresponding node. For each node matching value, perform an exponentiation operation with the natural constant e as the base to obtain the corresponding matching index value, and divide each matching index value by the sum of all matching indices to obtain the corresponding normalized response value. The average response value is obtained by summing the normalized response values ​​and dividing by the number of response values. Nodes with normalized response values ​​greater than the average response value are identified as conflict structural elements, thus obtaining a set of conflict structural elements.

6. The engineering design assistance method based on multi-source data fusion according to claim 1, characterized in that, Step four specifically involves: For each conflict structural element in the conflict structural element set, extract the corresponding normalized conflict intensity value and normalized response value; Add the normalized conflict intensity value to the normalized response value to obtain the boundary influence value of the corresponding conflict structural unit; The average boundary influence value is obtained by summing all the boundary influence values ​​in the set of conflicting structural elements and dividing by the number of conflicting structural elements. Conflicting structural elements whose boundary influence value is greater than the average boundary influence value are identified as conflicting structural elements that alter the design boundary. The conflicting structural element with the largest boundary influence value among the conflicting structural elements that change the design boundary is identified as the adjudication seed.

7. The engineering design assistance method based on multi-source data fusion according to claim 1, characterized in that, Step five specifically involves: Using the fragment corresponding to the adjudication seed as the reference fragment, for each fragment in the fragment set, calculate the difference vector between the node feature representation corresponding to the current fragment and the node feature representation corresponding to the reference fragment, and add the values ​​of each dimension in the difference vector to obtain the associated difference value. Add all the association difference values ​​together and divide by the number of fragments in the fragment set to get the average association difference value. Determine the fragments whose association difference value is less than or equal to the average association difference value as the association fragments that are close to the adjudication seed, and obtain the association fragment set. For each associated fragment in the associated fragment set, extract the corresponding node feature representation, and add the values ​​of each dimension in the node feature representation to obtain the fragment feature value; For any two associated fragments in the associated fragment set, calculate the difference between the feature values ​​of the two fragments, and perform an absolute value operation on the difference to obtain the fragment tension value; For each associated fragment, the tension values ​​of all fragments corresponding to the associated fragment are added together to obtain the overall tension value of the associated fragment. Divide the overall tension value of each associated fragment by the sum of all overall tension values ​​to obtain the corresponding normalized tension value. Based on all normalized tension values, construct the manifold topological tension field. The average tension value is obtained by summing all normalized tension values ​​and dividing by the number of associated fragments in the associated fragment set. Associated fragments with normalized tension values ​​less than the average tension value are identified as frozen fragments and frozen. Associated fragments with normalized tension values ​​greater than or equal to the average tension value are identified as unfrozen fragments.

8. The engineering design assistance method based on multi-source data fusion according to claim 1, characterized in that, Step six specifically involves: For each unfrozen fragment, extract the corresponding normalized tension value, and multiply the normalized tension value with the node feature representation corresponding to the unfrozen fragment dimension by dimension to obtain the fragment adjustment vector of the corresponding unfrozen fragment. The element-wise summation of all fragment adjustment vectors yields the fused adjustment vector; Determine the unfrozen geometry fragments and frozen geometry fragments; The fusion adjustment vector is added dimension by dimension to the node feature representation of each unfrozen geometric fragment to obtain the updated node feature representation of the unfrozen geometric fragment. Preserve the node feature representation of the frozen geometry fragments unchanged; The updated node feature representations of non-frozen geometric fragments are combined with the node feature representations of frozen geometric fragments to generate candidate design schemes.

9. The engineering design assistance method based on multi-source data fusion according to claim 1, characterized in that, Step seven specifically involves: Extract the node feature representations corresponding to all performance region fragments, add them element by element and divide by the number of performance region fragments to obtain a unified performance feature representation; The feature representations of each node in the candidate design scheme are subtracted dimension by dimension from the unified performance feature representation to obtain the corresponding feature difference vector; The absolute values ​​of each dimension in each feature difference vector are calculated and then summed to obtain the corresponding matching deviation value. The average matching deviation value is then calculated. The node features whose matching deviation value is less than or equal to the average matching deviation value in the candidate design scheme are retained to obtain the optimized design scheme.

10. The engineering design assistance method based on multi-source data fusion according to claim 1, characterized in that, Step eight specifically involves: Extract the node feature representations corresponding to the design intent expression fragments, add them element by element and divide by the number of design intent expression fragments to obtain a unified intent feature representation; The node feature representation corresponding to the optimized design scheme is multiplied dimension by dimension with the unified intent feature representation to obtain the product value of the corresponding dimension; The number of dimensions with product values ​​greater than or equal to zero is counted, and then divided by the number of dimensions represented by node features to obtain the consistency ratio value. The average consistency ratio value is then calculated. The node features in the optimized design scheme whose consistency ratio is greater than or equal to the average consistency ratio are retained and identified as auxiliary results for engineering design.

Images