Complex mechanical product demand tracing method based on demand management model

Abstract

The invention provides a complex mechanical product demand tracing method based on a demand management model, and the method comprises the steps: firstly analyzing a SysML demand model of a complex mechanical product, and constructing a heterogeneous graph structure with multiple attributes; constructing a heterogeneous graph neural network model based on an HGT architecture, and training the heterogeneous graph neural network model by adopting a training data set containing positive and negative samples and pseudo label data; verifying the correctness of the demand relationship in the SysML demand model of the complex mechanical product by using the trained heterogeneous graph neural network model, and identifying hidden demand association; and when a new demand or relationship is added, dynamically updating the heterogeneous graph neural network model through an incremental learning method. According to the method, automatic verification of the demand relationship and mining of potential association are realized, and the efficiency and reliability of demand management of complex mechanical products are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of model systems engineering technology, specifically to a method for tracing the requirements of complex mechanical products based on a requirements management model. Background Technology

[0002] Traditional R&D platforms for complex mechanical products typically integrate multiple tools and methods for requirements analysis and management. However, the lack of a unified requirements management model and automated requirements traceability mechanism makes it difficult to systematically verify and manage requirements relationships. This is especially true in the development of complex mechanical products involving multiple disciplines and levels, such as liquid rocket engines and helicopter transmission systems, where the dependencies and relationships between requirements are often quite complex. Traditional document-based requirements management methods are prone to omissions or errors in requirements and lack automated means to verify the correctness of requirements.

[0003] Model-Based Systems Engineering (MBSE) offers a new approach to solving the aforementioned problems. In recent years, R&D departments have gradually introduced MBSE methodologies and related tools (such as the SysML modeling language and DOORS requirements management tools) into several complex mechanical product development projects to achieve structured management and traceability of requirements. The applicant has also conducted related research, proposing "Requirements Management Methods, Systems, Equipment, and Media Based on a Unified Data Source Model" (Publication No. CN119004855A). However, existing requirements management models still have shortcomings in terms of automation and complexity handling, particularly lacking effective technical support for requirement relationship verification and potential correlation discovery. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention proposes a demand traceability method for complex mechanical products based on a demand management model. By parsing the SysML model to extract relevant entities and their relationships, a heterogeneous graph structure is constructed, and neural networks are used to verify demand relationships and mine potential associations, providing a more efficient and reliable solution for the development of complex mechanical products.

[0005] The technical solution of this invention is as follows:

[0006] A method for demand traceability of complex mechanical products based on a demand management model includes the following steps:

[0007] Step 1: Based on the MBSE method, establish a SysML requirement model for complex mechanical products according to the design requirements of complex mechanical products, parse the SysML requirement model file, and perform data cleaning and standardization to construct a heterogeneous graph structure with multiple attributes.

[0008] Step 2: Based on the heterogeneous graph structure established in Step 1, construct a heterogeneous graph neural network model based on the HGT architecture, and train the heterogeneous graph neural network model using a training dataset containing positive and negative samples and pseudo-label data to obtain a trained heterogeneous graph neural network model.

[0009] Step 3: Use the trained heterogeneous graph neural network model to verify the correctness of the demand relationships in the SysML demand model for complex mechanical products and identify hidden demand associations;

[0010] Step 4: When new requirements or relationships are added, the model is dynamically updated using incremental learning methods.

[0011] Furthermore, step 1 specifically includes the following steps:

[0012] Step 1.1: Analyze the SysML requirement model of complex mechanical products, extract the entities in the model as nodes and the relationships between entities as edges;

[0013] Step 1.2: Perform standardized cleaning processing on the parsed data, including completing missing attribute values, standardizing data units, and detecting outliers;

[0014] Step 1.3: Based on the extracted nodes and edges, construct a heterogeneous graph structure with multiple attributes.

[0015] Furthermore, in step 1, the entity includes requirements, components, subsystems, interfaces, and constraints, and the relationship includes dependency relationship, satisfaction relationship, derivation relationship, verification relationship, and implementation relationship.

[0016] Furthermore, in step 1, the heterogeneous graph structure is stored in the form of a directed multigraph, wherein the nodes contain type labels and detailed attributes. The type labels include requirements, components, subsystems, interfaces, and constraints, and the detailed attributes include priority, source information, and status. The edges contain the relationships between nodes and their attributes. The relationships include dependency, satisfaction, derivation, verification, and realization, and the edge attributes include relationship strength, directionality, degree of influence, and confidence.

[0017] Furthermore, step 2 specifically includes the following steps:

[0018] Step 2.1: Establish a training sample dataset, including positive and negative samples and pseudo-labeled samples; positive samples are the correct demand relationships extracted from the SysML demand model; negative samples are conflicting or known to be incorrect demand relationships constructed through random sampling; pseudo-labeled samples are generated by filtering unlabeled relationships with confidence using a pre-trained model.

[0019] Step 2.2: Configure the parameters of the heterogeneous graph neural network model; the heterogeneous graph neural network model adopts a multi-layer HGT convolutional structure, with multiple independent attention heads set in each layer;

[0020] Step 2.3: Train the heterogeneous graph neural network model, where the loss function combines the supervised loss of labeled data and the regularization term of unlabeled data;

[0021] Step 2.4: Evaluate the performance of the heterogeneous graph neural network model and make optimization adjustments.

[0022] Furthermore, in step 2.3, the supervision loss of the labeled data uses cross-entropy loss to measure prediction accuracy, and the learning process is constrained by a regularization term based on graph structure consistency.

[0023] Furthermore, in step 2.3, the performance of the validation sample set is continuously monitored during the training process, and an early stopping mechanism is triggered when the validation loss does not decrease for three consecutive rounds.

[0024] Furthermore, step 3 specifically includes the following steps:

[0025] Step 3.1: Using the trained heterogeneous graph neural network model, verify and predict the demand relationships in the SysML demand model for complex mechanical products. A demand traceability report will be generated based on the confidence level, containing the following: demand relationships with a confidence level greater than or equal to 0.9 are listed in the verified relationship list; demand relationships with a confidence level between 0.3 and 0.9 are listed in the potential relationship list; and demand relationships with a confidence level less than 0.3 are listed in the abnormal relationship warning list.

[0026] Step 3.2: Present the demand traceability report in graphical form. Use Graphviz to draw the relationship graph between nodes and edges. Based on different prediction results, different feature lines will be used to represent the results of the model demand traceability.

[0027] Furthermore, step 4 specifically includes the following steps:

[0028] Step 4.1: Obtain new requirements and relationship data from the SysML model, identify the added, modified and deleted elements in the SysML model, record the element ID, change type and specific content, and distinguish between substantive modifications and format adjustments through semantic analysis;

[0029] Step 4.2: Determine the minimum update range through change impact analysis and perform local fine-tuning on the heterogeneous graph neural network: only recalculate the network parameters corresponding to the affected elements and their associated nodes, freeze the network parameters of the unaffected parts, and fine-tune the parameter weights of the affected areas through the backpropagation algorithm to achieve rapid incremental updates while maintaining the overall stability of the model.

[0030] This invention also proposes a demand traceability model for complex mechanical products based on a demand management model, including a data preprocessing and graph construction module, a model training module, a demand analysis module, and a visualization module;

[0031] The data preprocessing and graph construction module extracts raw requirement data from the SysML model, cleans, transforms and preprocesses it, and then constructs a heterogeneous graph structure. Entities such as requirements and components are represented as nodes, and relationships are represented as edges. The nodes and edges are given corresponding attribute features to form a heterogeneous graph that can be used for neural network training.

[0032] The model training module uses a semi-supervised learning method, combining a small amount of labeled data and a large amount of unlabeled data, to train the heterogeneous graph neural network model, optimize model parameters, and improve the model's ability to verify demand relationships and discover hidden demand relationships.

[0033] The requirement analysis module analyzes and predicts the requirement relationships in the SysML model based on the trained model, evaluates the correctness of the requirement relationships, discovers potential missing requirement associations, and generates a requirement traceability report.

[0034] The visualization module presents the results of demand tracing in a graphical form, using Graphviz to display the demand relationship diagram.

[0035] Beneficial effects

[0036] This invention constructs a demand management model based on heterogeneous graph neural networks, which realizes automatic verification of demand relationships and mining of potential associations, significantly improving the efficiency and reliability of demand management for complex mechanical products. Furthermore, this invention dynamically updates the model through an incremental learning mechanism to adapt to constantly changing demands, solving the problem of lack of accuracy and real-time performance of traditional methods when dealing with large and complex products, and improving product development efficiency and quality.

[0037] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0038] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0039] Figure 1 Flowchart of a method for tracing the demand for complex mechanical products;

[0040] Figure 2 SysML model analysis diagram;

[0041] Figure 3 Flowchart of heterogeneous graph neural network training;

[0042] Figure 4 : Schematic diagram of semi-supervised joint training;

[0043] Figure 5 Example diagram of demand traceability visualization;

[0044] Figure 6 : Schematic diagram of incremental learning mechanism. Detailed Implementation

[0045] 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.

[0046] To address the challenges posed by the multi-level decomposition and cross-domain relationships inherent in large and complex mechanical products, which traditional tools struggle to comprehensively cover and effectively analyze, and the low automation and potential omissions in existing demand tracing tools like SysML and DOORS when tracing the demands of complex mechanical products, this embodiment proposes a demand tracing method for complex mechanical products based on a demand management model. This method utilizes a heterogeneous graph neural network to automatically verify and locate demand relationships within the SysML model, making it suitable for demand tracing of complex mechanical products. It significantly improves the automation level and reliability of demand tracing, which is of great importance for enhancing product development efficiency and ensuring product quality.

[0047] This method parses model elements using the SysML V2 API to construct a heterogeneous graph structure containing various types of nodes (requirements, components, etc.) and relational edges (dependencies, satisfactions, etc.). A heterogeneous graph neural network with an HGT architecture is employed, combined with semi-supervised learning for model training, utilizing cross-entropy loss and regularization terms to optimize parameters. The trained model is used to verify requirement relationships, identify hidden associations, and generate a visual graph using Graphviz (solid lines represent verified relationships, dashed lines represent potential associations). Furthermore, incremental learning allows for updating model data, enabling continuous optimization and complete traceability of requirement relationships.

[0048] This embodiment focuses on a helicopter transmission system and includes the following steps:

[0049] Step 1: Guided by the MBSE Harmony-SE methodology, use M-design to build a SysML requirement model for the helicopter transmission system based on its design requirements. Parse the SysML requirement model file using the SysML V2 API, perform data cleaning and standardization, construct a heterogeneous graph structure with multiple attributes, and create multi-level indexes for the constructed graph structure to support efficient subsequent queries and analysis.

[0050] This step is specifically as follows:

[0051] Step 1.1: Use the SysML V2 API to parse the input SysML requirement model of the helicopter transmission system, extracting entities as nodes and relationships between entities as edges. Entities include requirements, components, subsystems, interfaces, constraints, etc., and relationships include dependencies, fulfillments, derivations, verifications, implementations, etc. For example, in a helicopter transmission system, requirement entities include performance indicators and reliability requirements, while component entities include key components such as the main reducer and drive shaft.

[0052] Step 1.2: Perform standardization and cleaning on the parsed data. This mainly includes three operations: completing missing attribute values; standardizing units to unify parameter units; and detecting outliers and removing isolated nodes and duplicate edges. These preprocessing steps ensure the integrity and consistency of the data for subsequent modeling.

[0053] Step 1.3: After data cleaning, a heterogeneous graph structure with multiple attributes is constructed based on the extracted nodes and edges. This graph structure is stored in the form of a directed multigraph, where nodes contain type labels and detailed attributes. Type labels include requirements, components, subsystems, interfaces, constraints, etc., and detailed attributes include priority, source information, status, etc. Specifically, requirement nodes have attributes such as requirement ID, description, priority, and source information; component nodes have attributes such as component ID, type, material, and functional description; and subsystem nodes have attributes such as subsystem ID, constituent components, and their configuration information. Edges contain the relationships between nodes and their attributes. Relationships include dependency, satisfaction, derivation, verification, realization, etc. These relationship types can clearly represent the relationships between different nodes. By setting these relationship types, requirement tracing and verification can be effectively performed. Edge attributes include relationship strength, directionality, degree of influence, confidence, etc. Specifically, relationship strength quantifies the tightness of the relationship between requirements, helping to identify high-priority or high-impact requirement relationships in the SysML requirement model. Direction clarifies the dependency path between requirements, ensuring the logical clarity and accuracy of the tracing process.

[0054] Ultimately, the structured data in the model can be fully retrieved through API calls, and NetworkX and Pandas are used to organize and store it to support subsequent demand tracking; and multi-level indexes are built for the constructed graph structure to support efficient subsequent queries and analysis.

[0055] Step 2: Based on the heterogeneous graph structure established in Step 1, construct a heterogeneous graph neural network model based on the HGT architecture. Train the model using a training dataset containing positive and negative samples and pseudo-labeled data to obtain the trained heterogeneous graph neural network model. Joint modeling of labeled and unlabeled data is performed using a semi-supervised learning method to improve the model's generalization ability and prediction accuracy.

[0056] This step is specifically as follows:

[0057] Step 2.1: Establish a high-quality training sample dataset; specifically, it includes three types of samples: positive samples are the correct demand relationships extracted from the SysML demand model; negative samples are conflicting or known to be incorrect demand relationships constructed through random sampling; pseudo-labeled samples are generated by the pre-trained model to filter unlabeled relationships by confidence, with the confidence threshold set to 0.9, so as to make full use of the heterogeneous graph structure information.

[0058] Step 2.2: Configure the heterogeneous graph neural network model parameters. The heterogeneous graph neural network model adopts a multi-layer HGT convolutional structure, with multiple independent attention heads in each layer, and the hidden layer dimension is set to 256. The model input is a heterogeneous graph structure with attributes, and the output is the relationship prediction probability.

[0059] Step 2.3: Train the heterogeneous graph neural network model. The loss function combines the supervised loss of labeled data and the regularization term of unlabeled data to ensure that the model can effectively utilize the structural information of unlabeled data while fitting the labeled data. Specifically, the supervised loss of labeled data uses cross-entropy loss to measure prediction accuracy, while the model learning process is constrained by a regularization term based on graph structure consistency to prevent overfitting. The Adam optimizer is used to update parameters, with the initial learning rate set to 0.001, along with a dynamic learning rate decay strategy. The performance of the validation sample set is continuously monitored during training. When the validation loss does not decrease for three consecutive rounds, an early stopping mechanism is triggered to improve training efficiency. This training process supports GPU-accelerated training.

[0060] Step 2.4: Evaluate and optimize the performance of the heterogeneous graph neural network model; that is, calculate metrics such as precision and recall on independent test sets to analyze the model's accuracy in verifying known relationships and predicting hidden relationships. For the parts with inaccurate predictions, add specific datasets to train the model and fine-tune the parameters to ensure that the model's prediction performance meets the standards.

[0061] Step 3: After obtaining the trained heterogeneous graph neural network model, use the model to verify the correctness of the demand relationships in the SysML demand model of the helicopter transmission system and identify hidden demand associations. The heterogeneous graph neural network model captures the complex dependencies between nodes and edges, discovers potentially missed demand associations, and provides a confidence score for each prediction result to quantify its credibility.

[0062] This step is specifically as follows:

[0063] Step 3.1: Using the trained heterogeneous graph neural network model, verify and predict the demand relationships in the SysML demand model for the helicopter transmission system. An automatic demand traceability report will be generated based on the confidence level, including the following: demand relationships with a confidence level greater than or equal to 0.9 will be listed in the verified relationship list; demand relationships with a confidence level between 0.3 and 0.9 will be listed in the potential relationship list for user reference; and demand relationships with a confidence level less than 0.3 will be listed in the abnormal relationship warning list. This helps users fully understand the accuracy and completeness of the demand relationships.

[0064] Step 3.2: Present the requirement traceability report graphically, using Graphviz to visualize the relationship between nodes (e.g., requirements, components) and edges (e.g., satisfaction, dependencies). Different colored lines will be used to represent the model's requirement traceability results based on the prediction results. Validated requirement relationships are marked with different colors or shapes, while hidden requirement relationships are represented by dashed lines or special markers. High-quality PNG or PDF graphic files are generated for easy saving and further analysis. For example, requirement relationships validated as correct by the model are connected by green lines; requirement relationships with potential associations are connected by yellow dashed lines; and abnormal requirement relationships in the original SYSML are connected by red lines. Users can view these relationships between the interactive interface and the requirement relationships, facilitating analysis of the model results. In this embodiment, the model prediction results are as follows: Figure 5 As shown.

[0065] Step 4: Connect with the MBSE simulation platform and use incremental learning to dynamically update the model when new requirements or relationships are added to the platform, in order to maintain the effectiveness and reliability of the model. Specifically, extract the newly created SysML model file from the MBSE platform, and after data cleaning, integrate the node and edge information in the graph into the existing heterogeneous graph, feed it into the heterogeneous graph neural network, thereby triggering the fine-tuning of the neural network while updating the graph. The neural network adjusts some parameters through the backpropagation algorithm to complete the incremental learning of the model.

[0066] This step is specifically as follows:

[0067] Step 4.1: Periodically obtain new requirements and relationship data from the SysML model of the MBSE simulation platform, identify the added, modified and deleted elements in the SysML model, record the element ID, change type and specific content, and distinguish between substantive modifications and format adjustments through semantic analysis.

[0068] Step 4.2: Determine the minimum update range through change impact analysis and perform local fine-tuning on the heterogeneous graph neural network: only recalculate the network parameters corresponding to the affected elements and their associated nodes, freeze the network parameters of the unaffected parts, and fine-tune the parameter weights of the affected areas through the backpropagation algorithm. This achieves rapid incremental updates while maintaining the overall stability of the model. In this way, a small number of iterative training sessions are performed on the basis of the already trained model using new data, avoiding the resource consumption caused by training from scratch and ensuring that the model can quickly adapt to the ever-changing demand management environment.

[0069] Furthermore, this embodiment also proposes a demand traceability model for complex mechanical products based on a demand management model, including a data preprocessing and graph construction module, a model training module, a demand analysis module, and a visualization module, which work together to complete the demand traceability task; wherein:

[0070] The data preprocessing and graph construction module extracts raw requirement data from the SysML model, cleans, transforms and preprocesses it, and then constructs a heterogeneous graph structure. Entities such as requirements and components are represented as nodes, and relationships are represented as edges. The nodes and edges are given corresponding attribute features to form a heterogeneous graph that can be used for neural network training.

[0071] The model training module uses a semi-supervised learning method, combining a small amount of labeled data with a large amount of unlabeled data, to train the heterogeneous graph neural network model, optimize the model parameters, and improve the model's ability to verify demand relationships and discover hidden demand relationships.

[0072] The requirements analysis module analyzes and predicts the requirements relationships in the SysML model based on the trained model, evaluates the correctness of the requirements relationships, discovers potential missing requirements associations, and generates a requirements traceability report.

[0073] The visualization module presents the results of demand traceability in a graphical form, using Graphviz to display demand relationship diagrams, helping users intuitively understand the process and results of demand traceability and providing support for decision-making.

[0074] Through the collaborative work of the above modules, this invention can automate, intelligentize, and visualize the traceability of demand for complex mechanical products, improve the efficiency and reliability of demand management, reduce errors and rework in the development process, and thus improve product quality and R&D efficiency.

[0075] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A method for demand traceability of complex mechanical products based on a demand management model, characterized in that: Includes the following steps: Step 1: Based on the MBSE method, establish a SysML requirement model for complex mechanical products according to the design requirements of complex mechanical products, parse the SysML requirement model file, and perform data cleaning and standardization to construct a heterogeneous graph structure with multiple attributes. Step 2: Based on the heterogeneous graph structure established in Step 1, construct a heterogeneous graph neural network model based on the HGT architecture, and train the heterogeneous graph neural network model using a training dataset containing positive and negative samples and pseudo-label data to obtain a trained heterogeneous graph neural network model. Step 3: Use the trained heterogeneous graph neural network model to verify the correctness of the demand relationships in the SysML demand model for complex mechanical products and identify hidden demand associations; Step 4: When new requirements or relationships are added, the model is dynamically updated using incremental learning methods.

2. The method for demand traceability of complex mechanical products based on a demand management model according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1.1: Analyze the SysML requirement model of complex mechanical products, extract the entities in the model as nodes and the relationships between entities as edges; Step 1.2: Perform standardized cleaning processing on the parsed data, including completing missing attribute values, standardizing data units, and detecting outliers; Step 1.3: Based on the extracted nodes and edges, construct a heterogeneous graph structure with multiple attributes.

3. The method for demand traceability of complex mechanical products based on a demand management model according to claim 2, characterized in that: In step 1, the entities include requirements, components, subsystems, interfaces, and constraints, and the relationships include dependency relationships, satisfaction relationships, derivation relationships, verification relationships, and implementation relationships.

4. The method for demand traceability of complex mechanical products based on a demand management model according to claim 3, characterized in that: In step 1, the heterogeneous graph structure is stored in the form of a directed multigraph, where nodes contain type labels and detailed attributes. Type labels include requirements, components, subsystems, interfaces, and constraints, and detailed attributes include priority, source information, and state. Edges contain the relationships between nodes and their attributes. Relationships include dependency, satisfaction, derivation, verification, and realization, and edge attributes include relationship strength, directionality, degree of influence, and confidence.

5. The method for demand traceability of complex mechanical products based on a demand management model according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 2.1: Establish a training sample dataset, including positive and negative samples and pseudo-labeled samples; positive samples are the correct demand relationships extracted from the SysML demand model; negative samples are conflicting or known to be incorrect demand relationships constructed through random sampling; pseudo-labeled samples are generated by filtering unlabeled relationships with confidence using a pre-trained model. Step 2.2: Configure the parameters of the heterogeneous graph neural network model; the heterogeneous graph neural network model adopts a multi-layer HGT convolutional structure, with multiple independent attention heads set in each layer; Step 2.3: Train the heterogeneous graph neural network model, where the loss function combines the supervised loss of labeled data and the regularization term of unlabeled data; Step 2.4: Evaluate the performance of the heterogeneous graph neural network model and make optimization adjustments.

6. The method for demand traceability of complex mechanical products based on a demand management model according to claim 5, characterized in that: In step 2.3, the supervision loss of the labeled data uses cross-entropy loss to measure prediction accuracy, and the learning process is constrained by a regularization term based on graph structure consistency.

7. The method for demand traceability of complex mechanical products based on a demand management model according to claim 5, characterized in that: In step 2.3, the performance of the validation sample set is continuously monitored during the training process, and an early stopping mechanism is triggered when the validation loss does not decrease for three consecutive rounds.

8. The method for demand traceability of complex mechanical products based on a demand management model according to claim 1, characterized in that: Step 3 specifically includes the following steps: Step 3.1: Using the trained heterogeneous graph neural network model, verify and predict the demand relationships in the SysML demand model for complex mechanical products. A demand traceability report will be generated based on the confidence level, containing the following: demand relationships with a confidence level greater than or equal to 0.9 are listed in the verified relationship list; demand relationships with a confidence level between 0.3 and 0.9 are listed in the potential relationship list; and demand relationships with a confidence level less than 0.3 are listed in the abnormal relationship warning list. Step 3.2: Present the demand traceability report in graphical form. Use Graphviz to draw the relationship graph between nodes and edges. Based on different prediction results, different feature lines will be used to represent the results of the model demand traceability.

9. The method for demand traceability of complex mechanical products based on a demand management model according to claim 1, characterized in that: Step 4 specifically includes the following steps: Step 4.1: Obtain new requirements and relationship data from the SysML model, identify the added, modified and deleted elements in the SysML model, record the element ID, change type and specific content, and distinguish between substantive modifications and format adjustments through semantic analysis; Step 4.2: Determine the minimum update range through change impact analysis and perform local fine-tuning on the heterogeneous graph neural network: only recalculate the network parameters corresponding to the affected elements and their associated nodes, freeze the network parameters of the unaffected parts, and fine-tune the parameter weights of the affected areas through the backpropagation algorithm to achieve rapid incremental updates while maintaining the overall stability of the model.

10. A demand traceability model for complex mechanical products based on a demand management model, characterized in that: It includes modules for data preprocessing and graph construction, model training, requirements analysis, and visualization. The data preprocessing and graph construction module extracts raw requirement data from the SysML model, cleans, transforms and preprocesses it, and then constructs a heterogeneous graph structure. Entities such as requirements and components are represented as nodes, and relationships are represented as edges. The nodes and edges are given corresponding attribute features to form a heterogeneous graph that can be used for neural network training. The model training module uses a semi-supervised learning method, combining a small amount of labeled data and a large amount of unlabeled data, to train the heterogeneous graph neural network model, optimize model parameters, and improve the model's ability to verify demand relationships and discover hidden demand relationships. The requirement analysis module analyzes and predicts the requirement relationships in the SysML model based on the trained model, evaluates the correctness of the requirement relationships, discovers potential missing requirement associations, and generates a requirement traceability report. The visualization module presents the results of demand tracing in a graphical form, using Graphviz to display the demand relationship diagram.

Images