AI-based CAE post-processing result automatic interpretation and report generation method and system

Abstract

The invention provides an AI-based CAE post-processing result automatic interpretation and report generation method and system, and the method comprises the steps: S1, obtaining multi-format computer-aided engineering simulation result data through a system, and generating structured data through analysis and conversion; s2, automatically identifying key features in the structured data by adopting an artificial intelligence model in combination with a domain knowledge graph, generating a feature list and performing visual labeling; s3, on the basis of the domain knowledge graph, deducing physical causes and engineering influences of the key features, and generating an analysis conclusion; and S4, according to a preset template and a language model, generating a standardized report containing the key features and the analysis conclusion. According to the method, automation from data processing to report generation is realized, the manual intervention cost is remarkably reduced, the reliability and decision support capability of complex engineering problem analysis are improved, and the method is suitable for various scenes such as structural engineering and fluid engineering.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an AI-based method and system for automatically interpreting and generating reports of CAE post-processing results. Background Technology

[0002] Computer-aided engineering post-processing is a crucial step in the field of engineering simulation, widely applied in structural design, fluid analysis, and other scenarios, directly impacting the quality and efficiency of product development. This field provides engineers with decision-making support by analyzing simulation data, encompassing the entire process from extracting key information from complex data to generating professional reports.

[0003] However, current methods have significant limitations when processing complex engineering data, restricting analysis efficiency and the reliability of results. Existing methods typically rely on manual or semi-automated data processing and analysis workflows, making it difficult to handle the diversity of simulation data formats. For example, structural engineering may simultaneously involve mesh data from finite element analysis and sensor data from experimental tests; these data have significantly different formats and structures, making them difficult to process uniformly.

[0004] Furthermore, existing tools often lack effective utilization of geometric topological information when identifying key features, leading to inaccurate analysis results. This is especially true in complex models, where feature identification and interpretation frequently require significant manual intervention, which is time-consuming and error-prone. Among the many technical challenges, the efficient integration and utilization of geometric topological information has become a core challenge. Geometric topological information describes the spatial structure and relationships of engineering models and is crucial for understanding simulation results. For example, in bridge structural analysis, geometric topological information determines the stress distribution pattern, but existing methods struggle to integrate it with industry-standard thresholds, making it impossible to accurately distinguish which stress concentration areas are normal phenomena and which are potential risks. This lack of information integration renders the analysis results unreliable in engineering scenarios, impacting subsequent decision-making.

[0005] Therefore, how to efficiently integrate geometric topological information with industry-standard thresholds under diverse data formats, automatically identify key features, and generate analysis results that meet engineering needs has become a critical issue in the post-processing of complex engineering projects. This technical challenge is particularly prominent in real-world engineering scenarios. Taking bridge design as an example, engineers need to determine whether the stress in key parts of the beam exceeds the safety threshold from simulation data. However, due to the diverse data sources, such as mesh data output by finite element software and point cloud data collected by monitoring equipment, the inconsistent formats make direct comparison difficult. Existing tools often rely on engineers to manually filter and compare these data, which is time-consuming and prone to errors due to subjective judgment. More importantly, geometric topological information is not fully explored in this process. For example, the curved structure of a bridge or the spatial characteristics of its connection nodes directly affect the interpretation of stress distribution, but existing methods struggle to automatically combine these characteristics with industry standards to determine which areas require special attention. This lack of information integration not only increases the complexity of the analysis but may also lead to the neglect of key issues, such as the failure to promptly identify potential structural weaknesses.

[0006] Therefore, how to seamlessly integrate geometric topology information with industry standard thresholds in a diverse data context to automatically identify and analyze key features has become a critical issue that engineering post-processing urgently needs to address. Summary of the Invention

[0007] This invention provides an AI-based method for automatic interpretation and report generation of CAE post-processing results, mainly including: It can access CAE result files in multiple formats, extract core data such as stress, flow field or temperature through a format parser, and convert them into structured data format. The extracted data is then denoised. At the same time, it extracts the correlation features of physical quantities, geometric features and engineering standards, and embeds material thresholds and industry standard thresholds to form standardized data. The preprocessed data is analyzed using a deep learning model, which includes an improved convolutional neural network for structural stress, thermal field or flow field, and an attention mechanism to focus on high gradient regions. The model inference outputs a list of key features, including feature ID, location, value, deviation from threshold and risk level. The features are verified by a domain knowledge graph to eliminate misidentification, and visual annotations are automatically overlaid on the result cloud map or animation. Based on a domain knowledge graph, the knowledge graph contains entity associations of physical quantities, geometric features, material properties, engineering standards, and failure modes. It uses reasoning logic of feature matching, cause investigation, and impact assessment to automatically infer the physical causes and engineering impacts of key features. At the same time, it calls the historical case library for cross-validation and supports manual correction to update the knowledge graph. Load a predefined scenario-based report template, which includes modules for summary, feature analysis, pattern inference, engineering suggestions, and appendix. Employ a fusion scheme of domain terminology library and fine-tuned large language model to automatically generate professional natural language text, populate feature data, inference conclusions, and visualization charts, and finally generate a multi-format report.

[0008] This invention also relates to an AI-based system for automatic interpretation and report generation of CAE post-processing results, the system comprising: The result data access module is configured to access CAE result files in multiple formats and extract core data such as stress, flow field or temperature through a format parser and convert them into structured data formats. The data preprocessing module is configured to perform noise reduction on the extracted data and extract the correlation features of physical quantities, geometric features and engineering standards, and embed material thresholds and industry standard thresholds. The AI ​​feature recognition module is configured to use a deep learning model to analyze the preprocessed data. The deep learning model includes an improved convolutional neural network for structural stress, thermal field or flow field, and combines an attention mechanism to focus on high gradient regions, outputting a list of key features, including feature ID, location, value, deviation from threshold and risk level, and verifying and eliminating false recognition through a domain knowledge graph. The pattern inference module is configured to be based on a domain knowledge graph, which contains entity associations of physical quantities, geometric features, material properties, engineering standards and failure modes. It uses inference logic of feature matching, cause investigation and impact assessment to automatically infer the physical causes and engineering impacts of key features and call the historical case library for cross-validation. The report generation module is configured to load predefined scenario-based report templates, which include modules for summary, feature analysis, pattern inference, engineering suggestions, and appendices. It adopts a fusion scheme of domain terminology library and fine-tuned large language model to automatically generate professional natural language text, fill in feature data, inference conclusions, and visualization charts, and generate multi-format reports.

[0009] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses an AI-based method for automatic interpretation and report generation of CAE post-processing results. It proposes an intelligent solution to address business scenarios such as diverse data formats, difficulties in feature recognition, complex causal analysis, and low report generation efficiency in the analysis of complex engineering simulation results.

[0010] This invention analyzes multi-format data, transforming it into a structured matrix. Combining deep learning models and domain knowledge graphs, it automatically identifies key features, performs causal inference, and generates standardized reports. Its core innovation lies in embedding geometric topological information and industry-standard thresholds into feature recognition. By integrating attention mechanisms and physical constraints, it ensures the engineering rationality of the analysis results. Simultaneously, it utilizes natural language generation technology to output professional reports, significantly improving analysis efficiency and accuracy.

[0011] This invention automates the process from data processing to report generation, significantly reducing the cost of manual intervention, improving the reliability of complex engineering problem analysis and decision support capabilities, and is applicable to various scenarios such as structural engineering and fluid engineering. Attached Figure Description

[0012] Figure 1 This is a flowchart of the AI-based method for automatically interpreting and generating reports of CAE post-processing results according to the present invention. Detailed Implementation

[0013] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0014] like Figure 1 The AI-based method for automatically interpreting and generating reports of CAE post-processing results in this embodiment may specifically include: S101. Access CAE result files in multiple formats, extract core data such as stress, flow field or temperature through a format parser, and convert them into structured data formats. Perform noise reduction on the extracted data. At the same time, extract the correlation features of physical quantities, geometric features and engineering standards, and embed material thresholds and industry standard thresholds to form standardized data.

[0015] In response to the diverse data formats of CAE results, such as ANSYS's .odb, LS-DYNA's .d3plot, and Fluent's .cas, as well as their complex dimensions, a standardized data system is constructed to lay the foundation for AI analysis. Unified access and conversion of multi-format result data Supports input of results from mainstream CAE solvers, achieving unified conversion through a "format parser + data mapping table": Format parser: Extracts core data from different solver result files (e.g., extracts nodal stresses σ_x / σ_y / σ_z from .odb files, extracts element displacements Ux / Uy / Uz from .d3plot files, and extracts flow field velocities Vx / Vy / Vz from .cas files). Data mapping table: Converts unstructured data (such as animation frames) into structured matrices. For example, flow field results are converted into a four-dimensional matrix (T×N×3×3, where T is the time step and N is the number of nodes) of "time step-mesh node ID-velocity value-coordinate". The converted data is formatted as JSON-LD (associated data format), which facilitates subsequent AI model calls and knowledge association. The conversion success rate is ≥99%, and the conversion time for a 10GB result file is ≤5 minutes.

[0016] As another implementation, stress distribution, flow field characteristics, and temperature field are extracted from multi-format CAE result files using a parser and converted into structured data to obtain an initial dataset. The Z-score method is used to detect outliers in the initial dataset, marking data points that deviate from the mean to obtain an outlier set. If the outlier set is not empty, the K-nearest neighbor algorithm is used to replace the outliers with the neighborhood mean, generating a denoised dataset. If the outlier set is empty, the initial dataset is used as the denoised dataset. Physical quantities and geometric features are extracted from the denoised dataset to generate a feature vector set. Based on preset material thresholds and industry standards, the feature vector set is standardized to generate a standardized dataset.

[0017] S102. Analyze the preprocessed standardized data using a deep learning model. The deep learning model includes an improved convolutional neural network for structural stress, thermal field, or flow field, and combines an attention mechanism to focus on high gradient regions. The model inference outputs a list of key features, including feature ID, location, value, deviation from the threshold, and risk level. The features are verified through a domain knowledge graph to eliminate misidentification, and visual annotations are automatically overlaid on the result cloud map or animation.

[0018] By integrating deep learning models with domain knowledge graphs, we can achieve fully automated and high-precision identification of key features in various types of CAE results, such as structural stress, flow field, and thermal field. Construction of Feature Recognition AI Model System Design dedicated AI models for different result types, covering core analysis scenarios in CAE post-processing: Model inference process: 1. Input the preprocessed result data, such as stress cloud matrix and flow field vector data; 2. The model uses an attention mechanism to focus on high gradient regions, such as areas of color abrupt change in stress cloud maps, and extracts feature vectors. 3. Verify the rationality of features by combining domain knowledge graphs and eliminate misidentifications, such as marking "stress anomaly points caused by mesh distortion" as invalid features; 4. Output feature recognition results: including location coordinates, numerical values, feature type, and risk level: red = danger, yellow = warning, green = safe.

[0019] Quantification and visualization of feature recognition results Quantitative output: Generate a "feature identification list" which includes feature ID, type, location (global coordinates X / Y / Z), value (e.g., σ=420MPa, v=50m / s, T=150℃), deviation from standard threshold, and risk level; Visual annotation: Automatically overlay feature annotations in the result cloud map / animation—for example, marking the maximum stress point with a red circle and attaching a numerical label; marking stress concentration areas with a yellow dashed box and labeling "Area area = 50mm". 2 The average σ is 380 MPa. The annotation results support interactive viewing.

[0020] As another implementation method, preprocessed data is obtained. Features are extracted from the preprocessed data using ResNet50 to obtain an initial feature set. Feature IDs, locations, and values ​​are extracted from the initial feature set. The deviation of each feature value from a preset threshold is calculated, and the preliminary risk level of the feature is determined based on the deviation. The Neo4j graph database is loaded, and query matching domain rules are executed on the features with the preliminary risk level to exclude mismatched features, resulting in a validated feature list. Based on the validated feature list, structured data containing feature IDs, locations, values, threshold deviations, and risk levels is generated. If the risk level of a feature in the structured data exceeds a preset threshold, high-risk features are visually labeled on the resulting cloud map, resulting in a labeled cloud map. Using the labeled cloud map, a dynamic visualization animation is generated to display the spatiotemporal distribution of high-risk features, resulting in the final visualization output.

[0021] S103. Based on a domain knowledge graph, the knowledge graph contains entity associations of physical quantities, geometric features, material properties, engineering standards, and failure modes. It adopts reasoning logic of feature matching, cause investigation, and impact assessment to automatically infer the physical causes and engineering impacts of key features. At the same time, it calls the historical case library for cross-validation and supports manual correction to update the knowledge graph.

[0022] By constructing a knowledge graph in the CAE domain and combining it with feature recognition results, the physical causes and engineering impacts of key phenomena can be automatically inferred, leading to interpretable analytical conclusions. 3.1 Construction of Knowledge Graph in CAE Domain The knowledge graph covers five categories of entities and their relationships: physical quantities, geometric features, material properties, engineering standards, and failure modes. It is stored using the Neo4j graph database, and its core content includes: Multi-dimensional pattern inference logic The analysis conclusions are automatically generated using a three-step reasoning method: "feature matching → cause investigation → impact assessment". Step 1: Feature Matching – Associating the identified key features (such as “σ=420MPa>σ_y=350MPa”) with entities in the knowledge graph to locate possible association factors (such as geometric features: bolt hole R=0.5mm; material properties: 45 steel σ_y=350MPa). Step 2: Cause Investigation – Eliminating irrelevant factors based on reasoning rules to pinpoint the core cause: Example: The maximum stress of a car's lower control arm exceeds the standard (σ=420MPa). Related factors include "bolt hole R=0.5mm", "load F=10kN", and "material σ_y=350MPa". Through knowledge graph reasoning: "load F=10kN corresponds to the design value and does not exceed the standard", "material σ_y meets the requirements", "bolt hole R=0.5mm < recommended value 1mm, and in historical cases, R<1mm led to stress concentration exceeding 20%" → core cause: "the bolt hole radius is too small (R=0.5mm) leading to stress concentration". Step 3: Impact Assessment – ​​Inferring the impact of the cause on engineering performance, such as “stress concentration will lead to a shortened fatigue life of 800,000 cycles, which is lower than the design requirement of 1 million cycles, and there is a risk of early fracture.”

[0023] Verification and correction of inference results Cross-validation: Call the historical simulation case library and compare the consistency between the current inference result and similar cases. If the deviation is >5%, such as the current inference being "the rounded corners are too small" and 90% of similar cases being "the load exceeds the limit", then the inference process is retried. Manual intervention interface: Supports engineers to correct the inference results, such as adding new causes such as "assembly error". The corrected results are automatically updated to the knowledge graph to optimize the accuracy of subsequent inference.

[0024] As another implementation, entity association data of physical quantities, geometric features, material properties, engineering standards, and failure modes are obtained from the knowledge graph. Numerical values ​​are extracted from the entity association data and combined into vectors to obtain an initial feature set. The cosine similarity between the initial feature set and preset engineering standards is calculated to determine the feature combination with the highest similarity. If the similarity is higher than a preset threshold, case data related to the feature combination is obtained from a historical case database to obtain a validation feature set. Failure modes from the validation feature set are input into a Bayesian network to infer causal relationships and determine physical causes. Based on the physical causes, the degree of deviation from the engineering standards is calculated using the root mean square error formula to obtain a deviation assessment result, where the root mean square error is equal to the square root of the sum of the squares of the actual values ​​corresponding to the physical causes minus the squares of the engineering standard values, divided by the number of samples. Key features with deviations greater than a preset threshold are extracted from the deviation assessment result, and the entity association data in the knowledge graph is updated to obtain a second knowledge graph. Numerical values ​​are extracted from the second knowledge graph and combined into a second vector to obtain an optimized feature set.

[0025] S104. Load a predefined scenario-based report template, which includes modules for summary, feature analysis, pattern inference, engineering suggestions, and appendix. It adopts a fusion scheme of domain terminology library and fine-tuned large language model to automatically generate professional natural language text, and fill in feature data, inference conclusions, and visualization charts to finally generate a multi-format report.

[0026] Design customizable report templates that combine feature recognition and pattern inference results to automatically populate content and generate natural language reports that meet engineering standards. Report template system design Based on industry scenarios, standardized templates for structural strength, flow field, thermal field, and fatigue life design are provided. The template structure includes "Report Summary → Result Feature Analysis → Physical Law Inference → Engineering Recommendations → Appendix (Data Tables / Charts)". Core module functions: Template customization: Enterprises can adjust the module order, terminology (e.g., change "σ_max" to "maximum principal stress"), and chart styles according to their own standards (e.g., IATF16949 for the automotive industry). The customized template can be saved as an exclusive version for the enterprise.

[0027] Automatic generation of natural language content The report employs a fusion solution of "domain terminology database + large language model (LLM)" to ensure that the language is professional, fluent, and unambiguous. Domain terminology database: Includes 100,000+ CAE professional terms, such as "stress concentration factor Kt", "Reynolds number Re", and "heat flux density q", avoiding colloquial expressions such as "stress is too great" and replacing them with "stress value exceeds the material's yield strength by 20%"; LLM optimization: Based on pre-trained LLM (such as GPT-4 fine-tuning), input "feature data + inference conclusion", generate natural language text, and then verify and correct it through a terminology database. For example, "the temperature is very high" is corrected to "the temperature value of 150℃ exceeds the heat resistance limit of aluminum alloy 120℃ by 25%"; Content logic verification: Ensure that the logic of each module in the report is consistent. For example, the "stress exceeding the standard" in "Result Feature Analysis" corresponds to the "cause of excessively small rounded corners" in "Law Inference". The "Engineering Suggestions" proposes solutions for the causes.

[0028] Report format generation and export Format support: Automatically generates PDF / Word / HTML formats to meet different application scenarios. PDF is used for formal reports, Word is used for secondary editing, and HTML is used for web sharing. Automated output: The generation process requires no manual intervention and includes chart insertion, page number annotation, table of contents generation, and company logo embedding; Version management: The report is automatically associated with the corresponding simulation result version (e.g., "Report V1.0 associated with Result-V1.1"), supporting version tracking and comparison with historical reports.

[0029] As another implementation, a scenario-based template is loaded, and the structural definitions of the summary module, feature analysis module, pattern inference module, engineering suggestion module, and appendix module are extracted from the scenario-based template to obtain a first template framework. A first domain terminology database is obtained, and a pre-trained language model is fused from the first domain terminology database. The first domain terminology database and the pre-trained language model are combined to generate a first professional terminology embedding model, resulting in a first terminology enhancement model. First feature data is extracted from a multi-dimensional data source, and the first feature data is preprocessed to remove noisy data, resulting in a first cleaned feature set. If the completeness of the first cleaned feature set is higher than a preset threshold, a random forest algorithm is used to infer patterns from the first cleaned feature set to generate a first inference conclusion set. If the completeness is lower than the preset threshold, missing data is supplemented to obtain a first complete feature set, and a random forest algorithm is then used to infer patterns from the first complete feature set to generate a first inference conclusion set. Based on the first inference conclusion set, GPT-2 is used in conjunction with the first terminology enhancement model to generate first natural language text, which fills the feature analysis module and the pattern inference module, resulting in first modular text content. Key data is extracted from the first cleaning feature set to generate bar charts and line charts, which are then used to fill the visualization chart areas in the first template framework, resulting in the first enhanced chart content. Based on the first modular text content and the first enhanced chart content, the summary module, engineering suggestion module, and appendix module in the first template framework are filled to generate the first complete report content.

[0030] This invention provides an AI-based method for automatic interpretation and report generation of CAE post-processing results. This method achieves intelligent analysis and interpretation of complex engineering simulation results through steps such as multi-format data parsing, intelligent feature recognition, knowledge graph reasoning, and automatic report generation.

[0031] The methods in the embodiments of the present invention have been described above. The system in the embodiments of the present invention is described below, and the system includes: The result data access module is configured to access CAE result files in multiple formats and extract core data such as stress, flow field or temperature through a format parser and convert them into structured data formats. The data preprocessing module is configured to perform noise reduction on the extracted data and extract the correlation features of physical quantities, geometric features and engineering standards, and embed material thresholds and industry standard thresholds. The AI ​​feature recognition module is configured to use a deep learning model to analyze the preprocessed data. The deep learning model includes an improved convolutional neural network for structural stress, thermal field or flow field, and combines an attention mechanism to focus on high gradient regions, outputting a list of key features, including feature ID, location, value, deviation from threshold and risk level, and verifying and eliminating false recognition through a domain knowledge graph. The pattern inference module is configured to be based on a domain knowledge graph, which contains entity associations of physical quantities, geometric features, material properties, engineering standards and failure modes. It uses inference logic of feature matching, cause investigation and impact assessment to automatically infer the physical causes and engineering impacts of key features and call the historical case library for cross-validation. The report generation module is configured to load predefined scenario-based report templates, which include modules for summary, feature analysis, pattern inference, engineering suggestions, and appendix. It adopts a fusion scheme of domain terminology library and fine-tuned large language model to automatically generate professional natural language text, and fill in feature data, inference conclusions, and visualization charts to generate multi-format reports. The visualization and interaction module is configured to automatically overlay visual annotations on the result cloud map or animation, and provides a user interaction interface for previewing and correcting inference results.

[0032] The system serves as the hardware and software implementation platform for the aforementioned methods. It adopts a client-server (C / S) architecture and comprises six core functional modules, all of which collaborate through standardized APIs. Hardware deployment architecture: Server-side: 2 physical servers (primary and backup architecture, Intel Xeon 8375C CPU, 256GB DDR5, 4TB NVMe SSD, NVIDIA A100 GPU) to deploy AI models, knowledge graphs, and case libraries; Client: Engineer terminal (Windows 11 / macOS Ventura), supports result data import, feature viewing, report preview and export. Minimum client configuration: Intel i5-12400 CPU, 16GB DDR4, NVIDIA GTX1650 GPU.

[0033] Taking "fatigue simulation post-processing of lower control arm of automobile chassis" as an example, the implementation process of this invention is explained in detail to verify the feasibility and effectiveness of the technical solution: Input data and environment preparation CAE result data: .d3plot file output by LS-DYNAR12.0, containing fatigue simulation results of the lower swing arm (stress, displacement, fatigue life), mesh number = 200,000, time step = 1000; Basic data: lower control arm geometric model (STEP format, bolt hole R=0.5mm), material parameters (45 steel, σ_y=350MPa, fatigue life design requirement=1 million cycles), industry standard (GB / T3075-2008). Hardware environment: Server (Intel Xeon 8375C, 256GB DDR5, A100 GPU), Client (Intel i7-13700K, 32GB DDR5, RTX4070); Software environment: This invention system V1.0, PyTorch2.1 (AI model framework), Neo4j5.15 (knowledge graph), PDFBox2.0.32 (report generation).

[0034] Results data preprocessing Data access and conversion: Import the .d3plot file, the system automatically parses and extracts stress data (σ_x / σ_y / σ_z) and fatigue life data, and converts it into JSON-LD format, which takes 8 minutes; Data denoising: Identify and replace 12 abnormal stress points (e.g., a node with σ=1200MPa is replaced with the neighborhood mean of 380MPa), and the data fluctuation after denoising is 1.8%≤2%; Feature enhancement: Associate stress data with geometric model, label “nodes around bolt holes” and “nodes in rounded corner areas”, and embed standard thresholds of σ_y=350MPa for 45 steel and 1 million fatigue life.

[0035] AI-driven automatic feature recognition Model inference: The improved CNN model is used to analyze stress data, focusing on the high gradient region around the bolt hole; Feature output: Generate a list of features for recognition. Feature ID=1: Maximum stress point, location (X=120mm, Y=50mm, Z=3mm), σ=420MPa, exceeding σ_y=350MPa by 20%, risk level=red (dangerous); Feature ID=2: Stress concentration area, location (around bolt hole), area = 62mm² 2 Average σ = 380 MPa, risk level = red; Feature ID=3: Fatigue life extreme value zone, location (stress concentration area), N=800,000 cycles, 20% lower than design requirements, risk level=yellow (warning). Visual annotation: Mark the maximum stress point with a red circle (with the label "σ=420MPa") in the stress cloud map, and mark the stress concentration area with a yellow dashed box to generate an annotated cloud map.

[0036] Intelligent inference of physical laws Knowledge graph association: The features “σ=420MPa exceeding the standard” and “bolt hole area” are associated with the knowledge graph, and the associated factors are matched: bolt hole R=0.5mm, load F=10kN (design value), material σ_y=350MPa; Cause investigation: Excluding load factors: F=10kN=design value, not exceeding the limit; Excluding material factors: σ_y = 350MPa meets the requirements; Locking in geometric factors: Bolt hole R=0.5mm < Knowledge Graph recommended value of 1mm, 85% of similar features in historical cases have stress concentration due to R being too small; Impact assessment: It is inferred that "the excessively small fillet radius of the bolt holes leads to stress concentration, which shortens the fatigue life to 800,000 cycles, which is lower than the design requirement, and there is a risk of early fracture within 3 years"; Cross-validation: 100 cases of "excessive stress in bolt holes" were called from the case library. 92% of them were consistent with the current inference (cause R is too small), and the validation was successful.

[0037] Standardized report generation Template selection: Load the "Automotive Structural Fatigue Simulation Report" template; Content population: Abstract: "Fatigue simulation of the lower control arm of the automobile chassis shows that the maximum stress of 420MPa exceeds the standard by 20%, and the fatigue life of 800,000 cycles does not meet the design requirements. The core cause is that the fillet radius of the bolt hole is too small (R=0.5mm), and the risk level is dangerous." Results analysis: Insertion of feature recognition table, labeled stress cloud diagram, fatigue life curve; Pattern deduction: The causes of excessive stress (too small bolt hole radius) and the impact of shortened service life (risk of early fracture) are explained in detail. The engineering recommendation is: "Increasing the fillet radius of the bolt holes to 1mm would reduce the expected stress to 320MPa and increase the fatigue life to 1.1 million cycles. The optimization effect needs to be verified by simulation again." Format generation: Automatically generates a PDF report with embedded company logo, taking 3 minutes to generate; Results Verification: The engineer previewed the report and confirmed that the feature recognition was accurate (σ=420MPa, position deviation ≤0.1mm) and the conclusions were reasonable. No corrections were needed, and the report was directly exported for project reporting.

[0038] In terms of report visualization, the system employs advanced graphics rendering technology and user interface design concepts. The generated charts are not only technically accurate but also visually appealing and professional. The system automatically adjusts the report's color scheme, font selection, and layout design according to the company's visual identity system, ensuring consistency with the corporate image. Furthermore, the system supports multilingual report generation, capable of producing reports in different languages ​​based on user needs, meeting the requirements of international projects.

[0039] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for automatically interpreting and generating reports of CAE post-processing results, characterized in that, The method includes: It can access CAE result files in multiple formats, extract core data such as stress, flow field or temperature through a format parser, and convert them into structured data format. The extracted data is then denoised. At the same time, it extracts the correlation features of physical quantities, geometric features and engineering standards, and embeds material thresholds and industry standard thresholds to form standardized data. The preprocessed standardized data is analyzed using a deep learning model, which includes an improved convolutional neural network for structural stress, thermal field or flow field, and an attention mechanism to focus on high gradient regions. The model inference outputs a list of key features, including feature ID, location, value, deviation from threshold and risk level. The features are verified by a domain knowledge graph to eliminate misidentification, and visual annotations are automatically overlaid on the result cloud map or animation. Based on a domain knowledge graph, the knowledge graph contains entity associations of physical quantities, geometric features, material properties, engineering standards, and failure modes. It uses reasoning logic of feature matching, cause investigation, and impact assessment to automatically infer the physical causes and engineering impacts of key features. At the same time, it calls the historical case library for cross-validation and supports manual correction to update the knowledge graph. Load a predefined scenario-based report template, which includes modules for summary, feature analysis, pattern inference, engineering suggestions, and appendix. Employ a fusion scheme of domain terminology library and fine-tuned large language model to automatically generate professional natural language text, populate feature data, inference conclusions, and visualization charts, and finally generate a multi-format report.

2. The method for automatic interpretation and report generation of CAE post-processing results according to claim 1, characterized in that, The access to multi-format CAE result files involves extracting core data such as stress, flow field, or temperature through a format parser and converting it into a structured data format. Noise reduction is performed on the extracted data. Simultaneously, correlation features between physical quantities, geometric characteristics, and engineering standards are extracted, and material thresholds and industry standard thresholds are embedded to form standardized data, including: The stress distribution, flow field characteristics, and temperature field are extracted from multi-format CAE result files using a parser and converted into structured data to obtain the initial dataset. Detect outliers in the initial dataset, mark data points that deviate from the mean, and obtain a set of outliers; If the set of outliers is not empty, the K-nearest neighbor algorithm is used to replace the outliers with the mean of the neighborhood to generate a denoised dataset; if the set of outliers is empty, the initial dataset is used as the denoised dataset; physical quantities and geometric features are extracted from the denoised dataset to generate a set of feature vectors. Based on preset material thresholds and industry standards, the feature vector set is standardized to generate a standardized dataset.

3. The method for automatic interpretation and report generation of CAE post-processing results according to claim 1, characterized in that, The process involves analyzing the preprocessed standardized data using a deep learning model. This model includes an improved convolutional neural network targeting structural stress, thermal field, or flow field, and incorporates an attention mechanism to focus on high-gradient regions. The model inference outputs a list of key features, including feature ID, location, value, deviation from a threshold, and risk level. Features are validated using a domain knowledge graph to eliminate false identifications, and visual annotations are automatically overlaid on the resulting cloud map or animation. Obtain standardized data; Standardized data is input into ResNet50 for feature extraction to obtain an initial feature set; Extract feature ID, location, and value from the initial feature set; Calculate the deviation of each feature value from the preset threshold, and determine the preliminary risk level of the feature based on the deviation; Load the graph database, execute query matching domain rules on the features of the preliminary risk level, exclude non-matching features, and obtain a verified feature list; Based on the verified feature list, generate structured data containing feature ID, location, value, threshold deviation, and risk level; If the risk level of a feature in the structured data exceeds a preset threshold, the high-risk feature is visually labeled on the resulting cloud map to obtain the labeled cloud map.

4. The method for automatic interpretation and report generation of CAE post-processing results according to claim 1, characterized in that, The domain-based knowledge graph, which includes entity associations for physical quantities, geometric features, material properties, engineering standards, and failure modes, employs reasoning logic based on feature matching, cause investigation, and impact assessment to automatically infer the physical causes and engineering impacts of key features. Simultaneously, it utilizes a historical case library for cross-validation and supports manual correction to update the knowledge graph, including: Extract entity association data of physical quantities, geometric features, material properties, engineering standards, and failure modes from knowledge graphs; Numerical values ​​are extracted from entity-related data and combined into vectors to obtain an initial feature set; The cosine similarity between the initial feature set and the preset engineering standard is calculated to determine the feature combination with the highest similarity. If the similarity is higher than the preset threshold, case data related to the feature combination is obtained from the historical case library to obtain the verification feature set. The failure mode in the verification feature set is input into the Bayesian network to infer the causal relationship and determine the physical cause. Based on the physical causes, the deviation from the engineering standard is calculated using the root mean square error formula to obtain the deviation assessment result. The root mean square error is equal to the square root of the sum of squares of the actual value corresponding to the physical causes minus the engineering standard value, divided by the number of samples. Extract key features with deviations greater than a preset threshold from the deviation assessment results, update the entity association data in the knowledge graph, and obtain the second knowledge graph; The second knowledge graph is used to extract numerical values ​​and combine them into a second vector to obtain an optimized feature set.

5. The method for automatic interpretation and report generation of CAE post-processing results according to claim 1, characterized in that, The loading of a predefined scenario-based report template, which includes modules for summary, feature analysis, pattern inference, engineering suggestions, and appendices, employs a fusion scheme of a domain terminology library and a fine-tuned large language model to automatically generate professional natural language text, populate it with feature data, inference conclusions, and visualization charts, ultimately generating a multi-format report, including: Load the scenario-based template, extract the structural definitions of the summary module, feature analysis module, pattern inference module, engineering suggestion module, and appendix module from the scenario-based template, and obtain the first template framework; Obtain a first domain terminology database, fuse a pre-trained language model from the first domain terminology database, combine the first domain terminology database with the pre-trained language model to generate a first professional terminology embedding model, and obtain a first terminology enhancement model; First feature data is extracted from multidimensional data sources, and the first feature data is preprocessed to remove noisy data, resulting in a first cleaned feature set. If the completeness of the first cleaned feature set is higher than a preset threshold, the random forest algorithm is used to infer the pattern of the first cleaned feature set and generate the first inference conclusion set. If the completeness is lower than a preset threshold, the missing data is supplemented to obtain the first complete feature set. Then, the random forest algorithm is used to infer the pattern of the first complete feature set and generate the first inference conclusion set. Based on the first set of inference conclusions, GPT-2 is used in conjunction with the first terminology enhancement model to generate the first natural language text, which is then filled into the feature analysis module and the pattern inference module to obtain the first modular text content. Key data is extracted from the first cleansing feature set, and bar charts and line charts are generated to fill the visualization chart area in the first template frame, thus obtaining the first chart enhancement content. Based on the first modular text content and the first chart enhancement content, the summary module, engineering suggestion module, and appendix module in the first template framework are populated to generate the first complete report content.

6. An AI-based automatic interpretation and report generation system for CAE post-processing results, characterized in that: The system includes: The result data access module is configured to access CAE result files in multiple formats and extract core data such as stress, flow field or temperature through a format parser and convert them into structured data formats. The data preprocessing module is configured to perform noise reduction on the extracted data and extract the correlation features of physical quantities, geometric features and engineering standards, and embed material thresholds and industry standard thresholds. The AI ​​feature recognition module is configured to use a deep learning model to analyze the preprocessed data. The deep learning model includes an improved convolutional neural network for structural stress, thermal field or flow field, and combines an attention mechanism to focus on high gradient regions, outputting a list of key features, including feature ID, location, value, deviation from threshold and risk level, and verifying and eliminating false recognition through a domain knowledge graph. The pattern inference module is configured to be based on a domain knowledge graph, which contains entity associations of physical quantities, geometric features, material properties, engineering standards and failure modes. It uses inference logic of feature matching, cause investigation and impact assessment to automatically infer the physical causes and engineering impacts of key features and call the historical case library for cross-validation. The report generation module is configured to load predefined scenario-based report templates, which include modules for summary, feature analysis, pattern inference, engineering suggestions, and appendices. It adopts a fusion scheme of domain terminology library and fine-tuned large language model to automatically generate professional natural language text, fill in feature data, inference conclusions, and visualization charts, and generate multi-format reports.

7. The system as described in claim 6, characterized in that, The system also includes a visualization interaction module, configured to automatically overlay visual annotations on the result cloud map or animation, and provide a user interaction interface for previewing and correcting inference results.

Images