Software testing verification method, device and storage medium

Abstract

The application discloses a software test verification method, device and storage medium, relates to the technical field of information retrieval, and comprises the following steps: in response to a software test instruction, performing feature vectorization processing on multi-dimensional evaluation data to obtain a feature vector, and constructing an engineering semantic graph based on the feature vector; matching the engineering semantic graph according to core elements of computer-aided engineering test requirements, determining a retrieval weight and a similarity threshold corresponding to the software test, calculating the similarity based on the retrieval weight and the similarity threshold in combination with the engineering semantic graph, determining a candidate result, and generating a test scheme according to the candidate result and user requirements; and configuring a software and hardware test environment based on the test scheme, driving target software to perform test to obtain a test result. The application solves the problem of low efficiency of computer software test verification by means of retrieval enhancement combined with graph construction, effectively improves retrieval matching precision, enhances test scheme adaptability, and optimizes overall test operation quality.

Description

Technical Field

[0001] This application relates to the field of information retrieval technology, and in particular to a software testing and verification method, device, and storage medium. Background Technology

[0002] In computer-aided engineering software testing and verification and industrial R&D quality control scenarios, the ability to intelligently retrieve multimodal test data, accurately match semantics, and perform closed-loop verification throughout the entire process is directly related to the testing quality of computer-aided engineering software, the reliability and safety of industrial products, and the overall efficiency of industrial R&D.

[0003] In related technologies, computer-aided engineering software testing and verification is carried out by manually consulting materials and performing tests, using a product lifecycle management system for centralized storage and keyword retrieval, and employing one-way retrieval. This approach relies on manual debugging and static configuration, which is difficult to adapt to the complex scenarios of computer-aided engineering software testing, resulting in low efficiency in computer software testing and verification.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this application is to provide a software testing and verification method, device, and storage medium, aiming to solve the technical problem of low efficiency in computer software testing and verification.

[0006] To achieve the above objectives, this application proposes a software testing and verification method, the method comprising: In response to software testing instructions, feature vectorization is performed on multi-dimensional evaluation data to obtain feature vectors, and an engineering semantic graph is constructed based on the feature vectors. Based on the core elements of computer-aided engineering testing requirements, the engineering semantic graph is matched to determine the retrieval weight and similarity threshold corresponding to this software test; Based on the retrieval weight and the similarity threshold, similarity is calculated using the engineering semantic graph to determine candidate results, and a test plan is generated based on the candidate results and user requirements. Based on the aforementioned test plan, a hardware and software testing environment is configured and the target software is driven to perform automated testing, thereby obtaining the test results of this software test.

[0007] In one embodiment, in response to the software testing instruction, multimodal data of the entire process of computer-aided engineering software testing and verification is acquired, and the multimodal data is classified and organized to obtain the multidimensional evaluation data; According to data processing rules and data format, the multidimensional evaluation data is classified and processed to obtain a classification feature set; By using a domain-customized model, the classification feature set is mapped into a unified-dimensional feature vector according to cross-modal feature fusion rules, thus obtaining the feature vector; Based on the correlation between the feature vectors, the relationships between graph nodes and graph edges are determined, and the engineering semantic graph is constructed by combining the graph update mechanism.

[0008] In one embodiment, the core elements of the computer-aided engineering testing requirements are extracted to determine the product object, test type, and key parameters of the software test. The product object, test type, and key parameters of this software test are matched with the engineering semantic graph to determine the retrieval intent to which this software test belongs. Based on the search intent matching the preset intent parameter configuration table, the search weight and similarity threshold corresponding to this software test are determined.

[0009] In one embodiment, based on the retrieval weights and the engineering semantic graph, the feature vector similarity, semantic association similarity, and scene adaptation similarity between the computer-aided engineering testing requirements and the knowledge base data are calculated respectively. The feature vector similarity, semantic association similarity, and scene adaptation similarity are weighted and fused according to a preset similarity weight to obtain the comprehensive similarity of the candidate data; Based on the overall similarity of the candidate data, initial data with an overall similarity higher than the similarity threshold are selected; The initial data is scored based on a comprehensive quality assessment of average similarity, timeliness, and diversity. The initial data are then sorted in descending order according to the comprehensive quality score to obtain the candidate results. The test plan is obtained by analyzing the user requirements and the test elements of the candidate results according to the standard test procedure and adapting the parameters.

[0010] In one embodiment, based on a general semantic graph, the core terminology system, parameter constraint rules, and scenario classification standards in the field of computer-aided engineering testing are extracted to form a domain knowledge framework template; The multidimensional evaluation data is classified, labeled, and structured according to the domain knowledge framework template to generate a multimodal data set; Cross-modal feature extraction and fusion processing are performed on a standardized multimodal dataset to obtain a set of domain knowledge feature vectors that correspond to the dimensions of the domain knowledge framework template. The domain knowledge feature vector set is associated and mapped with the domain knowledge framework template, and the node attributes and edge relationships of the engineering semantic graph are supplemented to obtain the knowledge base containing the original data, feature vectors and semantic associations.

[0011] In one embodiment, the quality scores of the initial data in terms of average similarity, timeliness, and diversity are calculated according to the three-dimensional scoring rules, and the quality scores are correlated with the initial data to obtain a three-dimensional quality score. Based on the search intent corresponding to this software test, the dimensional weight coefficients of the three-dimensional quality score are adjusted to obtain a quality score weight configuration that adapts to the search intent. The three-dimensional quality scores of the initial data are weighted and fused according to the quality score weight configuration to obtain the comprehensive quality score corresponding to the initial data. The initial data is sorted in descending order of comprehensive quality score, and duplicate or redundant data is removed from the initial data to obtain the candidate results.

[0012] In one embodiment, the test cases, parameter configurations, and execution judgment criteria of the test plan are parsed and converted into test scripts adapted to the target software interface; Configure the hardware and software testing environment according to the environment requirements of the test script, and install the dependent components and drivers to obtain the target testing environment; In the target testing environment, the target software being tested is driven to perform a fully automated test according to the test script, and process data, result data and exception log data are collected during the test. The process data, result data, and exception log data are formatted and their integrity is verified. Invalid data is removed and missing fields are added to generate the result data of this software test.

[0013] In one embodiment, an evaluation dataset is generated based on the test results of the current software test, combined with user feedback data on the search results and the test plan. Based on the evaluation dataset, the quantitative error index and qualitative reasonableness index of the test results, as well as the matching score of the test scheme in the search results, are calculated respectively. The quantitative error index, the qualitative reasonableness index, and the matching score are then weighted and calculated to obtain the suitability score. When the suitability score is lower than the preset suitability threshold, the parameter optimization process is triggered, a quality loss function is constructed, and an adaptive learning rate strategy is used to adjust the retrieval weight and the similarity threshold to complete the iterative optimization of the matching parameters. Based on the effective multimodal data from this software test, cross-modal feature extraction is performed using the matching parameters to obtain engineering knowledge data. The engineering knowledge data is embedded into the knowledge base, the node attributes and edge relationships of the engineering semantic graph are updated, and the configuration parameters corresponding to the search intent are adjusted in conjunction with the updated data.

[0014] In addition, to achieve the above objectives, this application also proposes a software testing and verification device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the software testing and verification method described above.

[0015] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the software testing and verification method described above.

[0016] This application provides a software testing and verification method, which includes performing feature vectorization processing on multi-dimensional evaluation data in response to software testing instructions and constructing an engineering semantic graph; determining retrieval weights and similarity thresholds based on the core elements of test requirements matching the graph; performing multi-dimensional similarity calculations to screen candidate results based on the two parameters; generating an adapted test plan based on user requirements; and then deploying the test environment and executing automated testing to obtain test results according to the plan. This method solves the technical problems of weak correlation of traditional computer-aided engineering software test data, insufficient accuracy of retrieval and matching, poor adaptability of test plans, and low test execution efficiency. It effectively improves the semantic correlation mining capability of test data, retrieval and matching accuracy, rationality of test plans, and overall operational efficiency of automated software testing.

[0017] In summary, this application constructs an engineering semantic graph through feature vectorization, sets retrieval weights and similarity thresholds to match requirements, filters candidate results to generate test plans, and conducts automated testing. This solves the problem of low efficiency in computer software testing and verification, effectively improves retrieval matching accuracy, enhances the adaptability of test plans, and optimizes the overall quality of testing operations. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the first embodiment of the software testing and verification method of this application; Figure 2 This is a flowchart of the method in this application; Figure 3 This is a diagram illustrating the system structure of this application; Figure 4 This is a flowchart illustrating the sixth embodiment of the software testing and verification method of this application; Figure 5 This is a flowchart illustrating the eighth embodiment of the software testing and verification method of this application; Figure 6 This is a schematic diagram of the software testing and verification equipment used in this application.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0023] In related technologies, computer-aided engineering software testing and verification is carried out by manually consulting materials and performing tests, using a product lifecycle management system for centralized storage and keyword retrieval, and employing one-way retrieval. This approach relies on manual debugging and static configuration, which is difficult to adapt to the complex scenarios of computer-aided engineering software testing, resulting in low efficiency in computer software testing and verification.

[0024] This application provides a solution: First, in response to software testing instructions, feature vectorization is performed on multi-dimensional evaluation data to obtain feature vectors, and an engineering semantic graph is constructed based on the feature vectors. Then, according to the core elements of computer-aided engineering testing requirements, the engineering semantic graph is matched to determine the retrieval weight and similarity threshold corresponding to this software test. Next, based on the retrieval weight, the similarity threshold, and the engineering semantic graph, similarity calculation is performed to determine candidate results. A test plan is generated based on the candidate results and user requirements. Finally, based on the test plan, a software and hardware testing environment is configured and the target software is driven to perform automated testing to obtain the test results of this software test.

[0025] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or software testing and verification device capable of performing the above functions. The following description uses a software testing and verification device as an example to illustrate this embodiment and the subsequent embodiments.

[0026] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0027] This application provides a software testing and verification method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the software testing and verification method of this application.

[0028] In this embodiment, the software testing and verification method includes steps S10 to S40: Step S10: In response to the software testing command, the multidimensional evaluation data is processed by feature vectorization to obtain feature vectors, and an engineering semantic graph is constructed based on the feature vectors.

[0029] Software testing instructions are the core control commands that trigger computer-aided engineering (CAD) software testing and verification tasks. Multidimensional evaluation data encompasses raw data from the entire CAD testing process, including text, images, numerical values, and extended categories. Examples include test case documents, industry standards and specifications, simulation cloud maps, test curves, sensor data, and software version update information. Feature vectors are digital representations of the inherent attributes and engineering characteristics of various evaluation data within a unified dimension. Engineering semantic graphs are interconnected knowledge networks in the CAD domain built upon entity nodes and related edge relationships.

[0030] In this embodiment, the aforementioned software test instructions can be triggered in four ways. First, through routine periodic inspections, the system automatically issues test instructions according to a preset maintenance cycle, periodically completing standard simulation tests and verifications of predetermined machine models and structural components. Second, through abnormal operating condition alarms, when abnormal simulation calculations, component performance exceeding limits, or software runtime errors are detected, test instructions are immediately generated to conduct fault tracing tests. Third, through manual business initiation, operators select the test object, test type, and evaluation criteria through the interactive interface, and manually submit a request; the system then generates the corresponding test instructions. Fourth, through version iteration adaptation, after the computer-aided engineering software completes a version update and the testing tools complete adaptation modifications, the system automatically issues instructions to conduct compatibility and functional retesting.

[0031] After the test management module receives the software test instruction, it begins to retrieve the multi-dimensional evaluation data stored in the storage medium and performs feature vectorization processing on the raw data.

[0032] For example, the feature vectorization processing of multidimensional evaluation data includes three differentiated implementation paths. The first is a refined vectorization process categorized by type. Data queues are divided into four categories: text, image, numerical, and extended data. Corresponding domain-specific fine-tuning models are called for each category, sequentially completing text semantic extraction, image engineering feature extraction, numerical key parameter filtering, and extended data association annotation, generating corresponding dimension feature vectors one by one. These vectors are then aggregated to form a complete feature vector set. This method is specifically adapted to the inherent forms of different data, and feature extraction aligns with the actual judgment points in engineering, effectively avoiding the problem of feature distortion due to cross-type data mixing. The second is a unified vectorization process that integrates all data across all domains. This method abandons the boundaries between data categories and uniformly inputs all multidimensional evaluation data into a cross-modal fusion model. Through an attention mechanism, it automatically learns the correlation weights between different modalities of data and directly outputs a unified dimension of fused feature vectors. This method can capture the implicit correlations between cross-modal data, improving the completeness and consistency of feature representation. The third method, incremental dynamic vectorization, extracts features only from the newly added multi-dimensional evaluation data in this test, reusing feature vectors from previously processed data and dynamically expanding the feature vector set through an incremental update mechanism. This approach significantly reduces redundant computation and substantially improves the efficiency of feature vectorization processing.

[0033] Once the complete set of feature vectors is obtained, an engineering semantic graph can be constructed based on the following methods.

[0034] In one alternative approach, entity information is first extracted from the feature vector set to generate five types of nodes: terms, parameters, scenarios, data, and products. Then, based on the similarity calculation results between feature vectors, four types of edge relationships are established: semantic association, parameter constraints, scenario adaptation, and product association, ultimately forming a complete engineering semantic graph. This method features intuitive construction logic, low computational cost, fast execution speed, and is suitable for typical test scenarios with small data scales.

[0035] In another alternative approach, a pre-built general engineering semantic graph framework is first loaded. Then, the generated feature vector set is matched and aligned with the framework, adding new nodes and edge relationships. Simultaneously, the attribute information of existing nodes and the weight parameters of edge relationships are updated, ultimately forming a customized engineering semantic graph adapted to this test. This method can fully utilize existing knowledge resources, significantly shorten the graph construction time, and improve the completeness and accuracy of the graph.

[0036] In an exemplary scheme for determining an engineering semantic graph, a pre-built general engineering semantic graph framework is first loaded. This framework includes the core terminology system, parameter constraint rules, and scene classification standards of the computer-aided engineering testing field. Entity extraction is then performed, identifying five types of entities—terminology, parameters, scenes, data, and products—from the feature vector set, generating corresponding node objects. Next, node matching is performed, comparing the extracted nodes with nodes in the basic framework based on similarity, merging identical nodes, and adding new nodes. Then, edge relationship construction is performed, establishing four types of edge relationships between nodes—semantic association, parameter constraint, scene adaptation, and product association—based on the similarity calculation results between feature vectors and domain rules, and assigning corresponding weight parameters to each edge relationship. Finally, graph verification is performed, ensuring the accuracy and completeness of the graph through a combination of semantic consistency checks and manual review, generating the final engineering semantic graph.

[0037] It should be noted that in some special cases, no new multidimensional evaluation data was generated in this test, so the feature vectorization process was skipped and the existing engineering semantic graph was reused directly.

[0038] Step S20: Based on the core elements of the computer-aided engineering testing requirements, match the engineering semantic graph to determine the retrieval weight and similarity threshold corresponding to this software test.

[0039] The core elements of computer-aided engineering testing requirements are key information describing the objectives, objects, and requirements of the testing task. Examples include product type, test type, key parameters, accuracy requirements, and delivery deadline. Search weight is a quantitative parameter used to measure the importance of different modalities and features in similarity calculations. The similarity threshold is a critical value used to determine whether search results meet the requirements; only results with a comprehensive similarity not lower than this threshold are retained.

[0040] In this embodiment, the core elements of computer-aided engineering testing requirements are input into the requirements parsing module to extract the product object, test type, key parameters and accuracy requirements, forming a set of requirements features.

[0041] For example, the determination of retrieval weights and similarity thresholds includes three differentiated implementation paths. The first is the preset rule matching method, which, based on the test type and product type, retrieves the system's preset retrieval weight and similarity threshold configuration table and directly matches the corresponding parameter values. This method is logically simple and stable, with a fast response speed, and is suitable for conventional standardized testing scenarios. The second is the intent-driven dynamic generation method, which identifies the retrieval intent of this test based on a set of requirement features, including four categories: standard query, case retrieval, data comparison, and fault diagnosis. Then, it dynamically generates appropriate retrieval weights and similarity thresholds based on the retrieval intent. This method can accurately match the requirements of different testing tasks, improving the relevance and effectiveness of the retrieval results. The third is the historical case migration method, which retrieves the historical test case with the highest similarity to the requirements of this test, reuses the retrieval weights and similarity thresholds used in that case, and fine-tunes them according to the accuracy requirements of this test. This method can fully utilize historical experience, improving the rationality and reliability of parameter configuration.

[0042] In an exemplary scheme for determining search weights and similarity thresholds, the system's pre-set search intent classification model and parameter configuration rule base are first loaded. Then, requirement parsing is performed, inputting the core elements of the computer-aided engineering testing requirements into the search intent classification model to identify the search intent type corresponding to this test. Next, parameter matching is performed, retrieving the corresponding initial search weights and similarity thresholds from the parameter configuration rule base based on the identified search intent type. Then, parameter fine-tuning is performed, adjusting the initial parameters according to the accuracy requirements and delivery deadlines of this test; higher accuracy requirements result in higher similarity thresholds, and tighter delivery deadlines result in higher weights for data timeliness. Finally, parameter validation is performed to confirm that the adjusted parameters are within a reasonable range, generating the final search weights and similarity thresholds.

[0043] It should be noted that in some special cases, this test is a completely new type of test conducted for the first time, with no corresponding preset rules or historical cases. In such cases, the system's default general parameter configuration will be used, and optimization will be carried out based on the evaluation results after the test is completed.

[0044] Step S30: Based on the retrieval weight, the similarity threshold, and the engineering semantic graph, similarity calculation is performed to determine candidate results, and a test plan is generated based on the candidate results and user requirements.

[0045] Similarity calculation is a process used to quantify the degree of matching between requirements and knowledge base data. Candidate results are sets of knowledge base data with a comprehensive similarity score not lower than the similarity threshold. The test plan is a structured document guiding the execution of this software test, including test standards, test case selection, parameter configuration, process steps, and judgment criteria.

[0046] In this embodiment, the demand feature set, retrieval weight, similarity threshold and engineering semantic graph are input into the retrieval matching module to start the similarity calculation and candidate result screening process.

[0047] For example, the similarity calculation and candidate result determination methods include three differentiated implementation paths. The first is a multi-dimensional weighted fusion method, which calculates the feature vector similarity, semantic association similarity, and scenario adaptation similarity between the requirements and knowledge base data, respectively, and then performs weighted fusion according to retrieval weights to obtain a comprehensive similarity. Results with a comprehensive similarity not lower than the similarity threshold are selected as candidate results. This method can comprehensively evaluate the matching degree between data and requirements, improving the accuracy of retrieval results. The second is a semantic path reasoning method, which performs multi-hop semantic reasoning based on the engineering semantic graph to mine the implicit association paths between requirements and data. The comprehensive similarity is calculated based on path length and path weight, and candidate results that meet the requirements are selected. This method can discover relevant data that traditional literal matching cannot recognize, improving the comprehensiveness of retrieval results. The third is a hybrid generative method, which first uses the multi-dimensional weighted fusion method to select preliminary candidate results, then uses the semantic path reasoning method to supplement implicitly relevant results, and finally merges them to obtain the final set of candidate results. This method balances the accuracy and comprehensiveness of retrieval and is suitable for complex testing scenarios.

[0048] Once the candidate result set is obtained, a test plan can be generated based on the following methods.

[0049] In one possible approach, the test plan template with the highest overall similarity is selected from the candidate result set. Based on the core requirements of this test, the parameter configurations, process steps, and judgment criteria in the template are modified to generate the final test plan. This method is fast, logically simple, and suitable for conventional standardized testing scenarios.

[0050] In another alternative approach, reusable elements from multiple relevant test plans are extracted from the candidate result set, including test standards, test cases, parameter configurations, and process steps. These elements are then integrated and optimized according to the needs of this test to generate a customized test plan. This method fully leverages the advantages of multiple historical cases, improving the rationality and completeness of the test plan.

[0051] In an exemplary test plan, the candidate result set is first evaluated for quality and then sorted from highest to lowest overall quality score. Next, element extraction is performed, extracting reusable elements such as test criteria, test cases, parameter configurations, process steps, and judgment criteria from the top-ranked candidate results. Then, element integration is performed, integrating and optimizing the extracted elements based on the core requirements of this test, removing redundant content and supplementing missing information. Next, logical verification is performed, checking the logical consistency and parameter constraint compliance of the test plan based on the engineering semantic graph, and correcting any non-compliant content. Finally, the plan is formatted, organizing the content according to the standard test document format to generate the final structured test plan.

[0052] It should be noted that in some special cases, the candidate result set is empty, in which case a blank test plan containing the basic test framework and prompt information is generated, which is then manually supplemented and improved.

[0053] Step S40: Based on the test plan, configure the software and hardware test environment and drive the target software to perform automated testing to obtain the test results of this software test.

[0054] A hardware and software testing environment refers to the set of hardware devices and software environments required to run the target software and perform testing tasks. Automated testing refers to the process where a system automatically executes the testing process and collects test results without human intervention. Test results refer to all data and documents generated during the software testing process, including simulation results, process logs, exception information, and evaluation reports.

[0055] In this embodiment, the generated test plan is input into the test execution module to start the hardware and software environment configuration and automated testing process.

[0056] For example, the configuration methods for software and hardware testing environments include three differentiated implementation paths. The first is script-driven configuration, which automatically executes pre-written environment configuration scripts based on the environment requirements in the test plan, completing software installation, parameter settings, and hardware resource allocation. This method offers fast configuration speed, high repeatability, and is suitable for conventional standardized testing environments. The second is workflow orchestration configuration, which breaks down the environment configuration process into multiple independent atomic tasks. These tasks are executed in a specified order using workflow orchestration tools, supporting parallel execution and error retries. This method offers high flexibility and fault tolerance, and is suitable for complex multi-component testing environments. The third is adaptive scheduling configuration, which dynamically allocates hardware resources and software instances based on the resource requirements of the test plan and the current system resource status, automatically adjusting configuration parameters to adapt to different hardware environments. This method fully utilizes system resources, improving the efficiency and adaptability of environment configuration.

[0057] Once the software and hardware testing environment is configured, automated testing can be performed using the following methods.

[0058] In one alternative approach, the test plan is converted into a test script that the target software can recognize, driving the target software to execute a fully automated test process according to the script, and collecting test process and result data in real time. This method has simple execution logic, high stability, and is suitable for common automated testing scenarios.

[0059] In another alternative approach, a test execution plan is generated based on the test plan. The test execution engine then schedules multiple test tasks to run in parallel according to the plan, while monitoring the execution status of each task, automatically handling exceptions, and generating execution logs. This method improves test execution efficiency and is suitable for large-scale batch testing scenarios.

[0060] In an exemplary scheme for executing automated testing, the test plan is first loaded and its test cases, parameter configurations, execution order, and judgment criteria are parsed. Next, test script generation is performed, converting the parsed content into executable test scripts recognizable by the target software. Then, environment verification is performed to confirm that the hardware and software testing environment meets the testing requirements. Next, test execution is performed, driving the target software to perform fully automated testing according to the test script, monitoring the software's running status in real time, and synchronously collecting process data, result data, and exception log data. Finally, the execution results are processed, the collected test data is normalized and its completeness is verified, invalid data is removed and missing fields are added, generating a standardized dataset of the software test results.

[0061] It should be noted that in some special cases, if an exception occurs during the testing process that cannot be handled automatically, the automated test will be paused, an exception report will be generated, and manual handling will be notified.

[0062] Further, please refer to Figure 2 , Figure 2 This is a flowchart of the method in this application. After the software testing process begins, a Computer-Aided Engineering (CAE) test verification knowledge base is first constructed. Then, test requirements are analyzed and search intent is identified. Next, search matching and structured test plan generation are performed. Then, automated testing is executed and data collection is completed throughout the process. Subsequently, multi-dimensional quality assessment and parameter adaptive optimization are carried out. Finally, the knowledge base is dynamically updated and parameters are adjusted in linkage, and the process ends.

[0063] A CAE software testing and verification method based on the RAG methodology: This method includes six core steps: knowledge base construction and optimization, test requirement parsing and intent recognition, RAG retrieval and matching and test plan generation, test execution and data collection, result evaluation and parameter optimization, and dynamic knowledge base updating, forming a closed-loop optimization mechanism throughout the entire process, as detailed below: First, knowledge base construction and optimization: The knowledge base is the core data support for CAE software testing and verification. It needs to achieve structured storage of multimodal data, extraction of professional features, and construction of semantic associations, specifically including: Multimodal data acquisition: Acquiring multimodal data throughout the entire CAE software testing and verification process, including existing data and real-time updated data, specifically categorized as follows: Text-based data: test case documents, industry standard specifications (GB / T, ISO, etc.), simulation result analysis reports, fault logs, technical manuals, and expert experience documents; Image-based data: simulation cloud maps such as stress cloud maps, temperature cloud maps, and displacement cloud maps; test curves such as load-displacement curves and frequency-amplitude curves; operation diagrams of test equipment; real-life photos of failed components; and geometric model diagrams. Data types include: numerical file formats such as .dat / .csv / .odb simulation results, raw data collected by sensors such as vibration acceleration, temperature, and pressure data, test process records, and product design parameter data; New extended data includes: CAE software version update logs, testing tool compatibility instructions, and cross-domain testing related data.

[0064] Customized preprocessing for multimodal data: Based on the characteristics of different data types, customized preprocessing models adapted to the CAE domain are used to generate feature vectors and structured data with uniform dimensions or compatibility. For text-based data: First, terminology is labeled and entities are identified using a CAE domain terminology dictionary. Then, a BERT model finely tuned with a CAE domain corpus (containing more than 100,000 professional corpora covering core areas such as structure, heat, fluid, and electromagnetism) is used for semantic vectorization, generating a 512-dimensional semantic vector. At the same time, core parameters such as test standard number, product model, and parameter thresholds are extracted from the text along with semantic association information, providing a foundation for semantic graph construction.

[0065] Image-based data: Visual features are extracted using a ResNet-50 model finely tuned from CAE image samples (covering various simulation cloud maps, test curves, and engineering schematics, with a sample size of no less than 50,000 images), generating a 2048-dimensional feature vector. For simulation cloud maps, image segmentation techniques are additionally used to extract regional geometric features and numerical distribution features (such as the location of stress peak regions and temperature gradient distribution). For test curves, key engineering information is obtained through curve fitting and feature point extraction (such as peak points, inflection points, and extreme points), generating structured feature descriptions.

[0066] For data-related data: First, the high-dimensional numerical data is reduced in dimensionality using the PCA dimensionality reduction algorithm, retaining key engineering features such as maximum stress value, temperature peak, data distribution standard deviation, and peak occurrence time; then, a 512-dimensional feature vector is generated through mapping using a fully connected layer; at the same time, the association mapping relationship between the data and the test scenario and product type is established.

[0067] Engineering Semantic Graph Construction: Based on preprocessed multimodal data, an engineering semantic graph is constructed in the CAE domain to achieve structured organization and correlation mining of professional knowledge. The graph nodes include terminology nodes such as "modal analysis" and "mesh generation", parameter nodes such as "maximum stress" and "peak temperature", scenario nodes such as "vibration testing of automotive parts" and "strength testing of aerospace structures", data nodes such as "specific test cases" and "standard documents", and product nodes such as "motor housing" and "battery pack". Graph edge relationships include semantic associations (e.g., the association between "modal analysis" and "vibration test"), parameter constraints (e.g., the constraint relationship of "maximum stress" ≤ 350MPa), scenario adaptation (e.g., "battery pack fast charging test" adapts to "temperature cloud map" data), and product associations (e.g., the association between "motor housing" and "vibration test case"). Graph update mechanism: Supports the addition of new data nodes and the updating of edge relationships. Through a combination of semantic similarity calculation and manual review, the accuracy and completeness of the graph are ensured.

[0068] Knowledge base storage architecture: It adopts a hybrid storage architecture of "vector database + relational database + file storage": the vector database stores feature vectors of various types of data for fast similarity retrieval; the relational database stores structured data and supports structured queries; the file storage system stores raw multimodal files to ensure data traceability.

[0069] The second step is test requirement analysis and intent identification: This step aims to accurately understand user test requirements, clarify search intent and core needs, and provide a basis for subsequent search matching and test plan generation. Test requirement collection: Collect search input information from CAE test engineers. Input formats support natural language expressions (e.g., "Search for industry standards and similar cases for vibration testing of automotive motor housings"), parametric input (e.g., product type = motor housing, test type = vibration test, standard type = GB / T), and multimodal input (e.g., upload a screenshot of the fault + text description "Search for the troubleshooting solution corresponding to this fault").

[0070] Requirements preprocessing and core element extraction: Natural language requirements: The CAE domain BERT model, finely tuned in the multimodal data customized preprocessing of the step knowledge base construction and optimization, is used to convert the requirement text into a 512-dimensional semantic vector. At the same time, core elements are extracted through named entity recognition and keyword extraction algorithms, including product name, test type, data purpose, key parameters, version requirements, accuracy level, standard type, etc. Multimodal requirements: The uploaded image data is processed using a multimodal data customization preprocessing model built and optimized with a step-by-step knowledge base to extract feature vectors and structured features. These features are then fused with textual description information to generate a comprehensive requirement feature vector, from which core elements are extracted.

[0071] Search Intent Classification: User search intents are classified using an improved Support Vector Machine (SVM) multi-classification model (introducing an attention mechanism to improve classification accuracy). The search intents include the following four categories, each with a clearly defined core demand and applicable scenario: Standards query: The core requirement is to obtain the specific requirements, testing steps, and judgment criteria of industry standards and technical specifications. The applicable scenario is to query the standard basis in the initial stage of test plan development. Case search: The core requirement is to search for simulation results, failure cases, and success experiences of similar products and similar test scenarios. The applicable scenarios are test plan optimization and failure cause analysis. Data comparison: The core objective is to compare the differences between different versions of test cases, simulation results of different solutions, and requirements of different standards. The applicable scenarios are multi-solution selection and test result verification. Troubleshooting: The core requirement is to retrieve fault logs, troubleshooting methods, and solutions that match the current test fault symptoms, and the applicable scenario is fault handling during the test process.

[0072] Intent weight configuration: Configure the core demand weights according to the type of search intent. For example, for standard queries, the timeliness weight is 0.5, the accuracy weight is 0.4, and the diversity weight is 0.1; for case searches, the relevance weight is 0.5, the diversity weight is 0.3, and the timeliness weight is 0.2. This provides a basis for subsequent search parameter configuration.

[0073] Thirdly, RAG retrieval and matching, and test plan generation: Based on the requirements analysis results and knowledge base data, RAG technology is used to achieve accurate retrieval and intelligent test plan generation. The specific process is as follows: Initial matching parameter configuration: Based on the search intent classification results and intent weights, configure the initial matching parameters, including multimodal data weights, core matching dimension weights, and similarity thresholds. Specific preset values ​​are shown in Table 1, the initial matching parameter mapping table. Table 1: Initial Matching Parameter Mapping Table

[0074] Note: For average similarity weights, As a weight for timeliness, For diversity weights, and The sum of the weights for multimodal data is 1, adapting to the priority requirements of different intents regarding data types; simultaneously, in multi-dimensional similarity calculation, the feature vector similarity weights... Semantic association similarity weight Scene adaptation similarity weight ,and This should be consistent with the subsequent parameter descriptions.

[0075] Multi-dimensional similarity calculation: Combining feature vectors from the knowledge base with the engineering semantic graph, multi-dimensional similarity calculation is performed to improve matching accuracy. Feature vector similarity: The cosine similarity algorithm is used to calculate the similarity between the user demand feature vector and the knowledge base data feature vector. ; Semantic Relationship Similarity: Based on the engineering semantic graph, calculate the semantic relationship between core requirement elements and knowledge base data nodes. It is calculated using the shortest path length between nodes, the weight of associated edges, etc. Scenario Adaptation Similarity: Based on factors such as the test scenario and product type in the requirements, calculate the matching degree between the scenario and the knowledge base data. ; Overall similarity: according to the formula Calculate the overall similarity, where , , ,and .

[0076] Candidate result screening and ranking: Candidate results with a comprehensive similarity ≥ threshold T are screened, and duplicate and expired data are removed (screening is based on timeliness weight); then, the three-dimensional core matching dimension scores of the candidate results are calculated: Average similarity ( ): Through formula Calculation, where The number of candidate results. For user demand semantic vectors, For the first The feature vector of each candidate result The cosine similarity function; Timeliness: (as defined by the formula) Calculation, where For the current time, For the first The publication / generation time of each candidate result, The decay rate (valued between 0.1 and 0.3, determined experimentally based on typical update cycles for different types of data in the CAE field: According to the standard specification λ=0.3, the timeliness weight of data decreases to [value missing] after one year. When the test case λ=0.2, the weight of the timeliness of the data after 2 years decreases. When the simulation data λ=0.1, the weights remain unchanged after 3 years. It can accurately adapt to the timeliness requirements of different types of CAE data. Diversity: (via formula) Calculation, where For the first The feature vector of each candidate result For the first Feature vectors of candidate results; Overall score: according to the formula Calculate the overall score for each candidate result, sort them in descending order of overall score, and select the Top 10 candidate results.

[0077] Test plan generation: Integrate the Top 10 candidate results with user requirements, input them into a large CAE domain vertical model (fine-tuned based on a general large model using CAE professional corpus), and generate a structured test plan, including test standard basis, test case selection, test parameter configuration, test process steps, expected result judgment criteria, data collection requirements, etc.; Support users to manually adjust and confirm the test plan to form the final execution plan.

[0078] Fourth, test execution and data acquisition: Based on the generated test plan, achieve automated execution and full-process data acquisition for CAE software testing and verification. Test environment adaptation: Automatically detects CAE software version, test equipment status, and hardware resource configuration, and configures and adapts environment parameters according to test plan requirements to ensure that the test environment meets the execution conditions; supports interface docking and collaborative work with mainstream CAE software (such as ANSYS, ABAQUS, LS-DYNA).

[0079] Automated test execution: The test plan is converted into an executable script through the script generation module, and the test process is automatically executed by calling the CAE software interface and the test equipment control interface. It supports the breakpoint resume function. When an abnormal interruption occurs during the test, the interruption point can be resumed to avoid data loss and repeated testing.

[0080] Full-process data acquisition: Real-time acquisition of multimodal data during test execution, including test process data (such as real-time stress values ​​and temperature change data), test result data (such as final simulation cloud maps, test curves, and numerical result files), equipment operation data (such as voltage, current, and operating status of the test equipment), and fault information data (such as error messages and screenshots of fault phenomena that occur during the test). The data acquisition frequency can be dynamically adjusted according to the test type, and the acquisition frequency can reach up to 100Hz in high-precision test scenarios.

[0081] Fifth, result evaluation and parameter optimization: The test results and search matching quality are evaluated from multiple dimensions, and parameter optimization is achieved by combining user feedback, forming a closed-loop improvement: Test result evaluation: Quantitative assessment: Compare the test results with the expected results criteria, and calculate quantitative indicators such as error values ​​(e.g., stress value error, temperature peak error), pass rate, and consistency. Qualitative assessment: Based on engineering semantic graphs and expert experience rules, qualitative analysis is conducted on the rationality and effectiveness of test results, such as fault mode identification and performance evaluation. Comprehensive evaluation report: Generates a comprehensive evaluation report that includes quantitative indicators, qualitative analysis, data visualization charts (such as result comparison curves and error distribution histograms), and improvement suggestions.

[0082] Search match quality assessment: User feedback collection: Collect real-time feedback from users on search results and test plans, including click operations (number of clicks, click order), dwell time (duration for viewing a single result), relevance score (out of 1-5), plan satisfaction score (out of 1-5), and the extent of manual adjustments (such as the number of adjustment items and the proportion of adjusted content). Feedback Score Calculation: User feedback results are converted into a FeedbackScore using a range-based scoring system. The specific rules are as follows: Click count: 1-2 clicks = 1 point, 3-6 clicks = 2 points, ≥7 clicks = 3 points; Duration of stay: <10 seconds = 1 point, 10-30 seconds = 2 points, >30 seconds = 3 points; Relevance rating: 1-2 points = 1 point, 3-4 points = 2 points, 5 points = 3 points; Solution satisfaction rating: 1-2 points = 1 point, 3-4 points = 2 points, 5 points = 3 points; Manual adjustment range: 3 points for no adjustment, 2 points for minor adjustments (≤3 items), and 1 point for major adjustments (>3 items); Feedback Score The average of the scores for the five items above is rounded to two decimal places. Content quality score calculation: Calculate overall quality : ; Calculate the final content quality score: ,in This is the overall quality weight (value 0.7). The feedback score weight (with a value of 0.3) and The weighting is completely consistent with the claims and the wording of the supplementary protection points.

[0083] Scoring Judgment: A preset quality score adjustment threshold S (range 0.9-1.1, default 1.0) is used. If the current parameter configuration is well adapted, then the current parameter configuration is good; if This triggers the parameter optimization process.

[0084] Parameter optimization: Loss function construction: Constructing the loss function , where L is the difference between the quality score and the threshold, reflecting the adaptation deviation of the current parameter; Parameter update: A gradient-based heuristic adjustment strategy is adopted to calculate the loss function relative to the initial matching parameters ( Approximate gradient update for multimodal data weights According to the formula Update matching parameters; where The parameter set for the current iteration. For the updated parameter set, The learning rate (initial value 0.01-0.05, range 0.001-0.1). Adaptive learning rate adjustment: Dynamically optimize the learning rate based on the trend of the loss function after parameter adjustment. : If the loss function shows a decreasing trend over multiple consecutive iterations, then increase the learning rate. (Increase by 0.01 each time, with a maximum of 0.1), to accelerate parameter convergence; If L oscillates (the absolute value of the difference between two adjacent iterations > 0.02) or increases, then decrease the learning rate. (Decrease by 0.005 each time, with a minimum of 0.001) to improve parameter update stability; Secondary matching verification: Using the updated matching parameters, the process of calculating the multi-dimensional similarity between RAG retrieval matching and the generated test plan is repeated to verify the effect of parameter optimization and ensure... .

[0085] Sixth, dynamic updates to the knowledge base: Establish a real-time update mechanism for the knowledge base to ensure the timeliness, completeness, and accuracy of the data. New Data Monitoring: Real-time monitoring of new data in the CAE software testing and verification process, including newly released industry standards, new test cases, newly generated simulation results, newly recorded fault logs, newly supplemented expert experience, and manual adjustments to test plans; supports both automatic monitoring and manual uploading methods for adding new data.

[0086] New data preprocessing: The preprocessing process for new data automatically triggers steps for knowledge base construction and optimization, including customized preprocessing of multimodal data and engineering semantic graph construction. This process completes vectorization, structured feature extraction, semantic association construction, and feature label generation, ensuring that the new data conforms to the knowledge base storage specifications.

[0087] New data entry and graph update: The preprocessed new data is stored in the corresponding storage module, and the engineering semantic graph is updated to add new nodes and edge relationships. By combining similarity calculation and manual review, the consistency between the new data and the existing data is ensured to avoid duplication and redundancy.

[0088] Adaptive parameter adjustment: The initial matching parameters for the corresponding search intent are adaptively adjusted based on the new data type. If the new data is an industry standard, the timeliness weight β of the standard query intent is increased (e.g., from 0.5 to 0.6) and the text data weight (e.g., from 0.8 to 0.85). If the new data is a failed case, the image data weight of the case search intent is increased (e.g., from 0.5 to 0.55) and the diversity weight γ (e.g., from 0.3 to 0.35). If the new data is a test case version update, the similarity threshold T of the new data comparison intent is increased (e.g., from 0.8 to 0.83) and the average similarity weight α (e.g., from 0.6 to 0.65). If the new data is high-dimensional simulation data, the data class weights of all intents are increased, the PCA dimensionality reduction parameters are optimized, and the accuracy of high-dimensional data processing is improved.

[0089] Second Embodiment This embodiment provides an exemplary scheme for constructing an engineering semantic graph. In this example, firstly, in response to software testing instructions, multimodal data of the entire process of computer-aided engineering software testing and verification is acquired and classified to obtain multidimensional evaluation data. Then, the multidimensional evaluation data is classified and processed to obtain a classification feature set. Next, the classification feature set is mapped to a unified-dimensional feature vector through a domain-customized model. Finally, the graph node and edge relationships are determined based on the correlation between feature vectors, and the engineering semantic graph is constructed by combining the graph update mechanism. Step S10 includes steps A11 to A14: Step A11: In response to the software testing instruction, acquire multimodal data of the entire process of computer-aided engineering software testing and verification, and classify and organize the multimodal data to obtain the multidimensional evaluation data.

[0090] Step A12: According to the data processing rules and data format, classify and process the multidimensional evaluation data to obtain the classification feature set.

[0091] Step A13: Using a domain-customized model, the classification feature set is mapped into a unified-dimensional feature vector according to cross-modal feature fusion rules to obtain the feature vector.

[0092] Step A14: Based on the correlation between the feature vectors, determine the graph node and graph edge relationships, and construct the engineering semantic graph by combining the graph update mechanism.

[0093] Multimodal data refers to the collection of raw data with different representations and storage structures generated throughout the entire process of computer-aided engineering software testing and verification. It serves as the fundamental data source for constructing engineering semantic graphs. Examples include test case documents, industry standards and specifications, simulation cloud maps, test curves, sensor data, fault logs, and software version update information.

[0094] Multidimensional evaluation data is a structured collection of data with unified classification standards and data boundaries, obtained by classifying and organizing multimodal data. It serves as the direct input for feature extraction and processing. Examples include text-based evaluation data, image-based evaluation data, numerical evaluation data, and extended evaluation data.

[0095] A classification feature set is a set of features corresponding to different data categories obtained by extracting features from multidimensional evaluation data of different categories. It serves as the basic input for cross-modal feature fusion. Examples include text semantic feature sets, image engineering feature sets, numerical parameter feature sets, and extended association feature sets.

[0096] Domain-specific models are cross-modal feature fusion models that are fine-tuned specifically for the field of computer-aided engineering testing. They can accurately capture the semantic relationships and parameter constraints unique to the engineering domain. Examples include multi-classification fusion models based on improved support vector machines (SVM) and domain-fine-tuned cross-modal models based on Transformers.

[0097] In this example, when responding to software testing commands to acquire multimodal data to obtain multidimensional evaluation data, it can be done using a full synchronous acquisition method. Alternatively, it can be done using a streaming incremental acquisition method, receiving real-time multimodal data generated during the testing process while simultaneously classifying and organizing it according to preset classification rules to accumulate multidimensional evaluation data, thereby completing the acquisition and preprocessing of multimodal data.

[0098] After acquiring the multidimensional evaluation data, the classification feature extraction process is initiated. Following data processing rules and formats, targeted processing is performed on the text, image, numerical, and extended multidimensional evaluation data to extract features corresponding to each category. These features are then aggregated to obtain a classification feature set. This set is then input into a domain-customized model, where feature fusion and dimensional unification are performed according to cross-modal feature fusion rules, resulting in a unified-dimensional feature vector. Finally, based on the correlation strength between feature vectors, the corresponding graph nodes and graph edge relationships are determined. Combined with the graph update mechanism, the engineering semantic graph is constructed. This layered and progressive feature processing and association construction improves the accuracy and domain adaptability of the engineering semantic graph, avoiding semantic distortion caused by biases in feature extraction from general models.

[0099] For example, there are two methods for determining graph nodes and edge relationships based on the association between feature vectors. The first is a rule-based hierarchical association construction method. First, according to preset rules in the field of computer-aided engineering testing, feature vectors are divided into five categories: terms, parameters, scenarios, data, and products, generating five types of graph nodes. Then, the cosine similarity between the feature vectors corresponding to different nodes is calculated. When the similarity exceeds a preset threshold, an edge relationship is established between the corresponding nodes, and the weight of the edge relationship is determined based on the similarity value. After each pair of nodes completes the similarity calculation and edge relationship determination, the edge relationship is simultaneously added to the graph edge relationship set. After all node pairs have been compared and determined, all nodes and edge relationships are summarized, and the engineering semantic graph is constructed using a graph update mechanism. This method uses a domain rule-driven hierarchical node division and similarity threshold determination calculation logic. By using rule constraints that fit the actual engineering practice, it ensures accurate classification of graph nodes and that edge relationships conform to domain common sense, thus improving the engineering practicality of the graph.

[0100] The second method is a dynamic association construction method based on an attention mechanism. This method performs global feature distribution analysis on all feature vectors, automatically learns the association weights between different feature vectors through a self-attention mechanism, generates an association weight matrix, and then performs clustering based on this matrix. Feature vectors with association weights higher than a preset threshold are grouped into the same node cluster, with each cluster corresponding to a graph node. Graph edge relationships are established based on the average association weights between node clusters, and the strength of these edge relationships is adjusted in real time according to the dynamic changes in association weights. After all node clusters are divided and edge relationships are established, all nodes and edge relationships are summarized, and the engineering semantic graph is constructed using a graph update mechanism. This method employs a self-attention mechanism-driven dynamic association learning and node clustering computational logic. By automatically mining implicit associations between feature vectors, it can discover potential edge relationships that traditional rule-based methods cannot identify, thus improving the completeness and semantic richness of the graph.

[0101] Further, please refer to Figure 3 , Figure 3 This is a diagram illustrating the system structure of this application. First, the multimodal data acquisition unit of the knowledge base module collects data. After processing by the customized preprocessing unit, the engineering semantic graph construction unit constructs the engineering semantic graph, which is then stored through the hybrid storage unit. Next, the system enters the requirement processing and intent recognition module. The requirement acquisition unit collects test requirements, and after the requirement preprocessing and element extraction unit extracts core elements, the intent classification and weight configuration unit completes the intent classification and retrieval weight configuration. Then, the system enters the retrieval-augmented generation (RAG) and solution generation module. The multi-dimensional similarity calculation unit performs similarity calculations, and after candidate result filtering and sorting by the candidate result filtering and sorting unit, the test solution generation unit generates test solutions. Next, the system enters the test execution and data acquisition module. The test environment adaptation unit configures the test environment, and the automated script generation unit generates the test scripts. The test execution unit then executes the automated tests, while the multimodal data acquisition unit collects test data throughout the entire process. Following this, the system enters the quality assessment and parameter optimization module. The test result evaluation unit evaluates the test results, and combined with user feedback collected by the user feedback acquisition unit, the quality score calculation unit calculates the comprehensive quality score. The parameter optimization unit then performs adaptive parameter optimization. Finally, the system enters the dynamic update module. The new data monitoring unit monitors newly added data, which is then processed by the new data preprocessing unit. The knowledge base update unit updates the knowledge base, and the parameter adaptive adjustment unit adjusts the corresponding search parameters accordingly.

[0102] A CAE software testing and verification system based on the RAG method: This system corresponds one-to-one with the methods mentioned above, adopting a modular design. It includes six core modules: a knowledge base module, a requirement processing and intent recognition module, a RAG retrieval and solution generation module, a test execution and data acquisition module, a quality assessment and parameter optimization module, and a dynamic update module. Each module has a clearly defined function and smooth data flow, as detailed below: The first module is the knowledge base module, which is used to collect, preprocess, store, and construct semantic associations for CAE testing and verification of multimodal data. This includes: Multimodal Data Acquisition Unit: Used to collect multimodal data throughout the entire CAE software testing and verification process, supporting both automatic monitoring and manual uploading, covering various data types such as text, images, and data; Customized Preprocessing Unit: Includes text preprocessing subunits, image preprocessing subunits, and data preprocessing subunits, employing CAE-domain fine-tuning models and customized algorithms respectively to complete feature extraction and vectorization processing of various data types; Engineering Semantic Graph Construction Unit: Used to construct and update the CAE-domain engineering semantic graph, realizing semantic associations of elements such as terms, parameters, scenarios, data, and products; Hybrid Storage Unit: Adopts a hybrid architecture of "vector database + relational database + file storage" to achieve classified storage and efficient retrieval of feature vectors, structured data, and raw files.

[0103] The second module is the requirement processing and intent recognition module, which is used to process user testing requirements, accurately extract core elements, and identify search intents, including: The requirements gathering unit collects users' test requirements information and supports various input formats such as natural language, parameterization, and multimodality. The requirements preprocessing and feature extraction unit uses a finely tuned BERT model and feature extraction algorithm from the CAE domain to transform requirements into feature vectors and extract core elements such as product name, test type, and key parameters. The intent classification and weight configuration unit classifies search intents using an improved SVM multi-classification model and configures core request weights and initial matching parameters according to intent type.

[0104] Thirdly, the RAG search and solution generation module is used to achieve accurate search and intelligent test solution generation, including: Multi-dimensional similarity calculation unit: Combines feature vector similarity, semantic association similarity, and scene adaptation similarity to calculate comprehensive similarity; Candidate result filtering and sorting unit: Filters and sorts candidate results based on similarity threshold and comprehensive score, and outputs Top 10 results; Test plan generation unit: Integrates candidate results with user needs, inputs them into a large vertical model of CAE domain, generates structured test plans, and supports manual adjustment and confirmation.

[0105] Fourth is the test execution and data acquisition module, which is used to automate the execution of the test plan and collect data throughout the entire process, including: Test Environment Adaptation Unit: Automatically detects and configures test environment parameters such as CAE software, test equipment, and hardware resources to ensure environment compatibility; Automated Script Generation Unit: Converts structured test plans into executable scripts, supporting interface integration with mainstream CAE software; Test Execution Unit: Calls the interfaces between CAE software and test equipment to automatically execute the test process, supporting breakpoint resumption; Multimodal Data Acquisition Unit: Acquires multimodal data of the test process and results in real time, ensuring data integrity and timeliness.

[0106] Fifth is the quality assessment and parameter optimization module, used to evaluate the quality of test results and search matching, and to optimize parameters, including: User Feedback Collection Unit: Collects user feedback on search results and test plans, covering various types of feedback data such as clicks, dwell time, and ratings; Feedback Score Calculation Unit: Converts user feedback into feedback scores using an interval scoring system; Quality Score Calculation Unit: Calculates the final quality score by combining the overall quality score with the feedback score; Parameter Optimization Unit: Constructs a loss function and updates matching parameters using a gradient heuristic adjustment strategy to achieve parameter optimization; Test Result Evaluation Unit: Performs quantitative and qualitative evaluation of the test results and generates a comprehensive evaluation report.

[0107] Sixth is the dynamic update module, used to realize real-time updates of the knowledge base and parameters, including: New Data Monitoring Unit: Real-time monitoring of new data in the CAE testing and verification process, supporting automatic monitoring and manual upload triggering; New Data Preprocessing Unit: Automatically preprocesses and constructs semantic associations for new data; Knowledge Base Update Unit: Stores preprocessed new data into the knowledge base and updates the engineering semantic graph; Adaptive Parameter Adjustment Unit: Adaptively adjusts initial matching parameters according to the new data type to ensure the compatibility of parameters with data characteristics.

[0108] Third Embodiment This embodiment provides an exemplary scheme for determining retrieval weights and similarity thresholds. In this example, the core elements of computer-aided engineering testing requirements are first extracted to determine the product object, test type, and key parameters for this software test. Then, these elements are matched with the engineering semantic graph to determine the retrieval intent to which this software test belongs. Finally, based on the retrieval intent, a preset intent parameter configuration table is matched to accurately determine the retrieval weights and similarity thresholds suitable for this software test. Step S20 includes steps B11-B13: Step B11: Extract the core elements of the computer-aided engineering test requirements, and determine the product object, test type, and key parameters for this software test.

[0109] Step B12: Match the product object, test type, and key parameters of the software test with the engineering semantic graph to determine the retrieval intent to which the software test belongs.

[0110] Step B13: Based on the search intent matching the preset intent parameter configuration table, determine the search weight and similarity threshold corresponding to this software test.

[0111] The product object refers to the specific engineering product or structural component targeted in this software test, serving as the core anchor point for matching nodes in the engineering semantic graph. Examples include automobile bodies, aircraft engine blades, bridge steel structures, and electronic chip packages. The test type refers to the specific category of tests performed in this software test, determining the scope and focus of the retrieved data. Examples include static testing, fatigue testing, thermal analysis testing, fluid dynamics testing, and collision testing. Key parameters are the engineering performance indicators that this software test needs to focus on and verify, serving as important references for parameter configuration. Examples include maximum stress value, fatigue life, peak temperature, pressure distribution, and deformation.

[0112] Search intent is the core requirement of this software test for the knowledge base data, determining the overall configuration strategy for search weights and similarity thresholds. Examples include standard query intent, case search intent, data comparison intent, and troubleshooting intent. The intent parameter configuration table is a pre-stored standardized set of search weights and similarity thresholds corresponding to different search intents, serving as the direct basis for parameter matching.

[0113] In this example, when extracting the core elements of computer-aided engineering test requirements to determine the product object, test type, and key parameters, a full text parsing approach can be used. Alternatively, a structured form extraction method can be employed, directly reading the corresponding fields from the standardized test requirement form submitted by the user to quickly obtain the product object, test type, and key parameters, thus completing the extraction of the core elements.

[0114] After extracting the core elements, the search intent identification process is initiated. The product object, test type, and key parameters of this software test are matched with the engineering semantic graph to determine the search intent to which this software test belongs. Then, based on the identified search intent, a preset intent parameter configuration table is matched to obtain the corresponding initial search weight and similarity threshold. Fine-tuning is then performed according to the accuracy requirements and delivery deadline of this test to finally determine the search weight and similarity threshold corresponding to this software test. In this way, through layered and progressive requirement analysis and intent matching, the accuracy and adaptability of search parameter configuration are improved, avoiding search result deviations caused by improper parameter configuration.

[0115] For example, there are two methods for determining the retrieval intent by matching product objects, test types, and key parameters with nodes in the engineering semantic graph. The first method is an intent recognition method based on precise node matching. First, the extracted product objects, test types, and key parameters are precisely matched with product nodes, test type nodes, and parameter nodes in the engineering semantic graph to find fully matching nodes. Then, based on the associations of these nodes in the engineering semantic graph, predefined intent mapping rules are queried to directly determine the retrieval intent to which this software test belongs. After each element's node matching is completed, the candidate intent set is updated synchronously. Once all elements have been matched, the intent with the highest matching degree is selected from the candidate intent set as the final retrieval intent. This method uses a computational logic of precise string matching and predefined rule mapping. Strict node matching ensures the accuracy of intent recognition. The logic is simple and stable, with a fast response speed, and is suitable for conventional standardized testing scenarios.

[0116] The second method is an intent recognition method based on multi-hop semantic reasoning. This method uses the extracted product object, test type, and key parameters as starting nodes. It performs up to three levels of multi-hop semantic reasoning within the engineering semantic graph to mine all associated nodes and edge relationships related to these starting nodes, generating an associated subgraph. Then, it performs semantic analysis and feature extraction on the associated subgraph, calculates the semantic similarity between the associated subgraph and each preset search intent, and selects the search intent with the highest semantic similarity as the final search intent. This method employs the computational logic of multi-hop semantic reasoning and subgraph semantic analysis, enabling it to uncover implicit relationships between requirement elements, handle fuzzy requirements and complex test scenarios, and discover potential search intents that traditional exact matching cannot identify, thus improving the comprehensiveness and adaptability of intent recognition.

[0117] Fourth embodiment This embodiment provides an exemplary scheme for generating a structured test plan. In this example, the three-dimensional similarity between the requirements and knowledge base data is first calculated based on the retrieval weights and the engineering semantic graph. Then, a weighted fusion is performed to obtain a comprehensive similarity and to filter the initial data. Next, the initial data is sorted by multi-dimensional quality scoring to obtain candidate results. Finally, the test elements of the user requirements and candidate results are analyzed and parameter adaptation is completed, thereby ultimately generating a structured test plan adapted to this software test. Step S30 includes steps C11 to C15: Step C11: Based on the retrieval weights and the engineering semantic graph, calculate the feature vector similarity, semantic association similarity, and scene adaptation similarity between the computer-aided engineering testing requirements and the knowledge base data.

[0118] Step C12: The feature vector similarity, semantic association similarity, and scene adaptation similarity are weighted and fused according to preset similarity weights to obtain the comprehensive similarity of the candidate data.

[0119] Step C13: Based on the overall similarity of the candidate data, select initial data whose overall similarity is higher than the similarity threshold.

[0120] Step C14: Calculate a comprehensive quality score for the initial data based on average similarity, timeliness, and diversity, and sort the corresponding initial data in descending order according to the comprehensive quality score to obtain the candidate results.

[0121] Step C15: Analyze the test elements of the user requirements and the candidate results according to the standard test process and perform parameter adaptation to obtain the test plan.

[0122] Feature vector similarity is a core indicator for quantifying the degree of matching between requirements and the underlying features of knowledge base data, reflecting the consistency between the literal meaning of the data content and the feature level. Examples include text content similarity, image feature similarity, and numerical parameter similarity. Semantic association similarity is an indicator that quantifies the strength of semantic association between requirements and data based on engineering semantic graphs, reflecting the implicit logical connections between data. Examples include terminology synonymy association, parameter constraint association, and causal inference association. Scenario adaptation similarity is an indicator that quantifies the degree of matching between the requirement testing scenario and the knowledge base data application scenario, reflecting the applicability of the data in a specific testing scenario. Examples include product type adaptation, test type adaptation, and working condition adaptation. Comprehensive similarity is a quantified value of the overall matching degree obtained by weighted fusion of the three dimensions of similarity, and is the core basis for selecting initial data.

[0123] The initial data is a knowledge base dataset with a comprehensive similarity score exceeding a similarity threshold, serving as the foundation for subsequent quality scoring and ranking. The comprehensive quality score is a weighted calculation of the initial data's average similarity, timeliness, and diversity, resulting in a quantitative value that forms the core basis for ranking candidate results. Test elements are the core components of the test plan, serving as the basic units for parameter adaptation and plan integration. These include test standards, test cases, parameter configurations, process steps, and judgment criteria.

[0124] In this example, the three-dimensional similarity can be calculated using the feature vector method. Alternatively, the multi-hop reasoning capability of the engineering semantic graph can be leveraged to uncover deep semantic relationships between requirements and data, supplementing the calculation of semantic association similarity and scene adaptation similarity, thereby completing the calculation of the three-dimensional similarity.

[0125] After completing the three-dimensional similarity calculation, the comprehensive similarity calculation process is initiated. The similarities of the three dimensions are weighted and fused according to the preset similarity weights to obtain the comprehensive similarity of each candidate data. Then, the initial data with a comprehensive similarity higher than the similarity threshold are selected. The initial data are then given a comprehensive quality score based on average similarity, timeliness, and diversity, and sorted in descending order to obtain the candidate results. Finally, the test elements of user needs and candidate results are analyzed and parameters are adapted to generate a structured test plan. In this way, through multi-dimensional matching and quality screening, the accuracy and applicability of the test plan are improved, and the plan deviation caused by single-dimensional matching is avoided.

[0126] For example, there are two methods for generating structured test plans. The first is a hierarchical, progressive similarity calculation and plan generation method. First, based on retrieval weights, feature vector similarity, semantic association similarity, and scene adaptation similarity are calculated separately. After each dimension's similarity calculation is completed, preliminary screening is performed simultaneously to remove obviously irrelevant data. Then, the remaining data is weighted and fused to obtain a comprehensive similarity, thus selecting the initial data. Next, the initial data undergoes multi-dimensional quality scoring and ranking to obtain candidate results. Then, test elements are extracted from the top-ranked candidate results, integrated and parameter-adapted according to standard testing procedures, and the final test plan is generated. This method employs a hierarchical, progressive calculation and screening logic. By reducing subsequent computational load through upfront dimensional screening, it improves the speed of plan generation while ensuring the accuracy of the plan and adapting it to conventional standardized testing scenarios.

[0127] The second method is a semantic path-driven candidate result mining and solution generation approach. Starting with the core elements of the requirement, multi-hop semantic reasoning is performed within the engineering semantic graph to mine all relevant semantic paths. Three-dimensional similarity is calculated based on the length and weight of the semantic paths to obtain a comprehensive similarity score and filter initial data. Next, the initial data is quality-scored and ranked based on the association strength of the semantic paths to obtain candidate results. Then, test elements from the candidate results are integrated according to the logical relationships of the semantic paths, automatically filling in missing test steps and adapting parameters to generate logically rigorous test solutions. This method uses semantic path-driven reasoning and integration logic, enabling the mining of relevant data that traditional similarity calculations cannot identify. Simultaneously, it automatically constructs the logical framework of the test solution based on semantic associations, improving the completeness and logical rigor of the solution and adapting to complex testing scenarios.

[0128] Fifth Embodiment This embodiment provides an exemplary scheme for constructing a CAE test and verification knowledge base. In this example, firstly, core terminology, parameter constraint rules, and scenario classification standards in the field of computer-aided engineering testing are extracted based on a general semantic graph to form a domain knowledge framework template. Then, multi-dimensional evaluation data is classified, labeled, and structured according to the framework template to generate a multimodal dataset. Next, cross-modal feature extraction and fusion processing are performed on the standardized multimodal dataset to obtain a domain knowledge feature vector set for the corresponding dimension. Finally, the feature vector set is associated and mapped with the framework template to supplement the engineering semantic graph, thereby ultimately constructing a CAE test and verification knowledge base containing original data, feature vectors, and semantic associations. Before step C11, steps D11~D14 are also included: Step D11: Based on the general semantic graph, extract the core terminology system, parameter constraint rules, and scenario classification standards in the field of computer-aided engineering testing to form a domain knowledge framework template.

[0129] Step D12: Classify, label, and structure the multidimensional evaluation data according to the domain knowledge framework template to generate a multimodal data set.

[0130] Step D13 involves performing cross-modal feature extraction and fusion processing on the standardized multimodal data set to obtain a set of domain knowledge feature vectors that correspond to the dimensions of the domain knowledge framework template.

[0131] Step D14: Associate and map the domain knowledge feature vector set with the domain knowledge framework template, supplement the node attributes and edge relationships of the engineering semantic graph, and obtain the knowledge base containing the original data, feature vectors and semantic associations.

[0132] A general semantic graph is a large-scale knowledge network encompassing universal human knowledge and conceptual associations, serving as a fundamental source for extracting domain-specific knowledge. Examples include general Chinese semantic graphs, general engineering and technical knowledge graphs, and semantic graphs of academic literature. A core terminology system is a standardized set of professional terms in the field of computer-aided engineering testing, representing the basic unit for expressing domain knowledge. Examples include stress, strain, fatigue life, thermal conductivity, fluid viscosity, modal frequency, and yield strength.

[0133] Parameter constraint rules are a set of rules describing the interrelationships between different parameters in computer-aided engineering testing. They are crucial for ensuring the rationality of test results and the effectiveness of engineering projects. Examples include allowable stress constraints, upper temperature limits, load range constraints, mesh accuracy constraints, and time step constraints. Scene classification standards are standardized rules for classifying computer-aided engineering testing scenarios. They are the foundation for data classification, organization, and precise retrieval. Examples include classification by product type, by test type, by operating conditions, by industry standards, and by verification level.

[0134] The domain knowledge framework template is a structured knowledge framework built upon a core terminology system, parameter constraint rules, and scenario classification standards. It serves as a unified standard for data annotation, feature extraction, and knowledge organization. The multimodal dataset is a multidimensional evaluation dataset categorized, annotated, and structured according to the domain knowledge framework template, possessing a unified data format, classification standards, and metadata specifications. Cross-modal feature extraction and fusion processing is the process of extracting and fusing the unique features of different modalities into a unified-dimensional feature vector, a key technology for achieving cross-modal data semantic association and unified retrieval. The domain knowledge feature vector set is a set of feature vectors with dimensions completely consistent with the domain knowledge framework template; each feature vector corresponds to a domain knowledge entity, attribute, or relationship.

[0135] A knowledge base is an integrated knowledge storage system that includes raw data, feature vectors, and semantic associations. It is the core foundation for retrieval-enhanced software testing methods.

[0136] In this example, when extracting domain knowledge based on a general semantic graph to form a domain knowledge framework template, a full knowledge scan approach can be used. Alternatively, domain experts can manually review the automatically extracted core terminology, parameter constraint rules, and scenario classification standards to verify, correct, and supplement the template, ensuring its accuracy, completeness, and engineering practicality, thereby completing the construction of the domain knowledge framework template.

[0137] After constructing the domain knowledge framework template, a data standardization preprocessing workflow is initiated. Multidimensional evaluation data is classified, labeled, and structured according to the framework template to generate a standardized multimodal dataset. Next, cross-modal feature extraction and fusion processing are performed on the multimodal dataset to obtain a set of domain knowledge feature vectors consistent with the dimensions of the framework template. Finally, the feature vector set is precisely associated and mapped with the domain knowledge framework template to supplement the node attributes and edge relationships of the engineering semantic graph, thus constructing a complete knowledge base. This standardized knowledge framework and cross-modal fusion technology improve the consistency, semantic richness, and retrieval efficiency of the knowledge base, avoiding retrieval biases caused by inconsistent data formats and semantic gaps.

[0138] For example, there are two ways to construct a knowledge base. The first is a rule-based standardized batch construction method. First, based on a general semantic graph and preset domain rules, the core terminology system, parameter constraint rules, and scenario classification standards of the computer-aided engineering testing domain are automatically extracted to generate a standardized domain knowledge framework template. Then, all historical multi-dimensional evaluation data are batch-classified, labeled, and structured according to the framework template to generate a unified format multimodal dataset. Next, batch cross-modal feature extraction and fusion processing are performed on the multimodal dataset to generate a domain knowledge feature vector set. Finally, the feature vector set is mapped one-to-one with the domain knowledge framework template to supplement the node and edge relationships of the engineering semantic graph, completing the initial construction of the knowledge base. This method uses rule-driven batch processing logic, unifies data format and feature dimensions through a standardized framework template, ensures the consistency and standardization of the knowledge base, has a fast construction speed, strong repeatability, and is suitable for initial knowledge base construction scenarios with small data scales.

[0139] The second approach is a dynamic iterative construction method based on incremental learning. First, an initial domain knowledge framework template and basic knowledge base are built. Then, newly added multidimensional evaluation data in the system are monitored in real time. The new data is incrementally classified, labeled, and structured according to the current domain knowledge framework template, generating an incremental multimodal data subset. Next, incremental cross-modal feature extraction and fusion processing are performed on the incremental multimodal data subset to generate an incremental domain knowledge feature vector subset. Finally, the incremental feature vector subset is mapped to the domain knowledge framework template, incrementally supplementing the node attributes and edge relationships of the engineering semantic graph. Simultaneously, the domain knowledge framework template is automatically iteratively optimized based on the distribution characteristics of the new data, achieving dynamic expansion and continuous optimization of the knowledge base. This method employs incremental learning and dynamic iteration, enabling real-time processing of new data, avoiding redundant calculations, and continuously improving the domain knowledge system, enhancing the timeliness and adaptability of the knowledge base, and adapting to long-term operation scenarios with continuously growing data.

[0140] Sixth Embodiment This embodiment provides an exemplary scheme for generating candidate results. In this example, the average similarity, timeliness, and diversity quality scores of the initial data are first calculated according to the three-dimensional scoring rules and correlated to obtain a three-dimensional quality score. Then, the dimensional weight coefficients of the three-dimensional quality score are adjusted according to the search intent corresponding to this software test to obtain an appropriate quality score weight configuration. Next, the three-dimensional quality scores are weighted and fused according to the weight configuration to obtain a comprehensive quality score. Finally, the initial data is sorted in descending order of comprehensive quality score and duplicate and redundant data are removed, thereby finally obtaining candidate results that accurately fit this software test. Please refer to... Figure 4 , Figure 4 This is a flowchart illustrating the sixth embodiment of the software testing and verification method of this application. Step C14 includes steps E11 to E14: Step E11: Calculate the quality scores of the average similarity, timeliness, and diversity of the initial data according to the three-dimensional scoring rules, and associate the quality scores with the initial data to obtain the three-dimensional quality score.

[0141] Step E12: Based on the search intent corresponding to this software test, adjust the dimension weight coefficients of the three-dimensional quality score to obtain a quality score weight configuration that adapts to the search intent.

[0142] Step E13: Perform a weighted fusion calculation on the three-dimensional quality score of the initial data according to the quality score weight configuration to obtain the comprehensive quality score corresponding to the initial data.

[0143] Step E14: Sort the initial data in descending order of comprehensive quality score, and remove duplicate and redundant data from the initial data to obtain the candidate results.

[0144] The three-dimensional scoring rules are a standardized set of calculation rules used to quantitatively evaluate the quality of initial data. They include independent calculation methods and scoring standards for three dimensions: average similarity, timeliness, and diversity. Examples include cosine similarity calculation rules, time decay function calculation rules, and content difference calculation rules. The quality scoring weight configuration is a set of weight coefficients for the three quality dimensions adjusted according to the search intent, with the sum of all dimension weight coefficients being 1. Examples include standard query intent weight configuration, case retrieval intent weight configuration, and troubleshooting intent weight configuration. Duplicate and redundant data refers to initial data that is highly similar or completely identical in content, including completely duplicated data, partially duplicated data, and derived duplicate data. Examples include different versions of the same test case, different expressions of the same parameter, and multiple test data under the same working condition.

[0145] In this example, when calculating the three-dimensional quality score of the initial data, the similarity calculation method can be used as a reference. Alternatively, by introducing a time decay factor and a content difference algorithm, the timeliness score and diversity score of the initial data can be calculated separately to complete the calculation of the three-dimensional quality score.

[0146] After calculating the three-dimensional quality score, the weight configuration adjustment process is initiated. Based on the search intent corresponding to this software test, the weight coefficients of the three dimensions of average similarity, timeliness, and diversity are dynamically adjusted to obtain an appropriate quality score weight configuration. Then, the three-dimensional quality scores are weighted and fused according to the weight configuration to obtain the comprehensive quality score corresponding to each initial data. Finally, the initial data are sorted in descending order of comprehensive quality score from high to low, and duplicate and redundant data are removed to obtain the final candidate results. In this way, the accuracy and comprehensiveness of the candidate results are improved through intent-driven dynamic weight adjustment and multi-dimensional quality assessment, avoiding the result bias caused by fixed weights.

[0147] For example, there are two ways to generate candidate results. The first is a batch scoring and sorting method based on pre-configured intent. First, the system loads a pre-set quality score weight configuration table corresponding to different search intents, and directly matches and obtains the corresponding weight configuration according to the search intent of this test. Then, the average similarity, timeliness, and diversity quality scores of all initial data are calculated in batches according to the three-dimensional scoring rules to generate a three-dimensional quality score set. Next, the three-dimensional quality scores of all initial data are batch-weighted and fused according to the matched weight configuration to obtain a comprehensive quality score set. Then, the initial data are sorted in descending order of comprehensive quality score from high to low. Finally, by combining hash comparison and similarity threshold judgment, duplicate and redundant data in the sorted data are removed to obtain the final candidate results. This method uses pre-configured weights and batch processing logic, which is fast and efficient, and can quickly generate candidate results, making it suitable for conventional standardized testing scenarios with high response speed requirements.

[0148] The second method is a dynamic weight adjustment ranking method based on feedback learning. First, an initial quality score weight configuration is obtained based on the search intent of the test. Then, the three-dimensional quality scores of all initial data are calculated and preliminarily weighted and fused to obtain an initial comprehensive quality score, generating a preliminary ranking result. Next, real-time user feedback data on the preliminary ranking result is collected, including click actions, dwell time, relevance scores, etc. A loss function is constructed based on the feedback data, and the weight coefficients of the three quality dimensions are dynamically adjusted using a gradient descent algorithm to obtain an optimized quality score weight configuration. Then, the comprehensive quality score of all initial data is recalculated according to the optimized weight configuration and sorted in descending order. Finally, duplicate and redundant data are removed to obtain the final candidate results. This method uses the logic of feedback learning and dynamic weight adjustment, which can continuously optimize the ranking results according to actual user needs, improve user satisfaction and engineering practicality of the candidate results, and adapt to complex and ever-changing test scenarios and personalized needs.

[0149] Seventh Embodiment This embodiment provides an exemplary scheme for executing automated testing and generating standardized test results. In this example, the test cases, parameter configurations, and execution criteria of the test scheme are first parsed and converted into test scripts adapted to the target software interface. Then, the hardware and software testing environment is configured according to the environment requirements of the test scripts, and dependent components and drivers are installed to obtain the target testing environment. Next, the target software is driven to perform a full-process automated test according to the test scripts in the target testing environment, and all data is collected. Finally, the collected data is normalized and its integrity is verified, invalid data is removed, and missing fields are supplemented, thereby ultimately generating standardized software test result data. Step S40 includes steps F11~F14: Step F11: Parse the test cases, parameter configurations, and execution judgment criteria of the test plan, and convert them into test scripts adapted to the target software interface.

[0150] Step F12: Configure the hardware and software testing environment according to the environment requirements of the test script, and install the dependent components and drivers to obtain the target testing environment.

[0151] Step F13: In the target test environment, drive the target software of this software test to perform a fully automated test according to the test script, and collect process data, result data and exception log data during the test.

[0152] Step F14: Normalize the format and verify the integrity of the process data, the result data and the exception log data, remove invalid data and fill in missing fields to generate the result data of this software test.

[0153] Test cases are sets of specific test scenarios and operational steps defined in a test plan to verify specific functions or performance of the target software. Examples include static loading test cases, fatigue cycle test cases, heat conduction analysis test cases, and fluid dynamics simulation test cases. Parameter configurations are sets of various parameters defined in a test plan that control the execution process of the target software. Examples include mesh generation parameters, load application parameters, boundary condition parameters, solver parameters, time step parameters, and convergence accuracy parameters. Execution judgment criteria are sets of standards and rules defined in a test plan to determine whether test results are acceptable. Examples include maximum stress threshold, maximum deformation threshold, fatigue life threshold, peak temperature threshold, and convergence residual threshold. Test scripts are computer program code containing a complete test process and parameter configuration that can be recognized and executed by the target software. Examples include APDL scripts, Python scripts, Tcl scripts, and macro command scripts.

[0154] A hardware and software testing environment is a collection of hardware devices and software environments required to run the target software and execute test tasks. Examples include high-performance computing servers, graphics workstations, operating systems, CAE software, database systems, and network environments. Dependent components are various auxiliary software components and library files required for the target software to run and execute tests. Examples include mathematical calculation libraries, graphics rendering libraries, parallel computing libraries, data parsing libraries, and communication protocol libraries. Drivers are software programs used to control the normal operation of hardware devices. Examples include graphics card drivers, network card drivers, storage device drivers, and peripheral device drivers. The target testing environment is a configured and verified hardware and software testing environment that fully meets the requirements of the test scripts; it is the foundational platform for executing automated tests.

[0155] The target software is the object of this software test, i.e., computer-aided engineering software that requires functional or performance verification. Examples include finite element analysis software, computational fluid dynamics software, multiphysics coupling analysis software, and structural optimization design software. Process data is data generated during test execution that reflects the test execution status and intermediate results. Examples include iteration steps, residual changes, computation time, resource utilization, and task execution progress. Result data is the final data generated after test execution that reflects the output results of the target software. Examples include stress contour maps, deformation contour maps, temperature field distribution, pressure field distribution, fatigue life curves, and displacement-time curves. Anomaly log data is log information generated during test execution that reflects software operational anomalies and errors. Examples include error messages, warning messages, crash logs, memory leak logs, and deadlock logs. Format normalization is the process of converting raw data of different formats and structures into a unified format and structure. Examples include converting result files from different software into a unified JSON format and converting contour maps of different resolutions into a unified image specification.

[0156] Integrity verification is the process of checking whether the collected data is complete and whether there are any missing data. For example, it checks whether all test cases have been executed completely, whether all result fields for each test case are complete, and whether exception logs are fully recorded. Invalid data refers to data that does not meet test requirements, contains errors, or is meaningless. Examples include non-convergent calculation results, abnormal data caused by hardware failures, redundant data collected repeatedly, and corrupted file data. Missing fields refer to fields that are missing from the collected data but should be present according to the standard. Examples include missing test timestamps, missing tester information, missing software version information, and missing environment configuration information. The result data is a standardized set of test data after format normalization and integrity verification, serving as the basis for subsequent quality assessment and knowledge base updates.

[0157] In this example, when parsing the test plan to generate the test script, it can be done through text parsing. Alternatively, it can be done by calling the official API interface provided by the target software, directly mapping the parameter configurations and execution steps in the test plan to the corresponding API call sequence, thereby generating a directly executable test script.

[0158] After generating the test script, the environment configuration process is initiated. The hardware and software testing environment is configured according to the test script's requirements, necessary dependencies and drivers are installed, and the target testing environment is obtained after environment consistency verification. Then, the target software is driven to perform fully automated testing according to the test script within the target testing environment, collecting process data, result data, and exception log data in real time. Finally, the collected data is normalized and its completeness is verified, invalid data is removed, and missing fields are supplemented to generate standardized software test result data. This automated test execution and data processing improves testing efficiency and result consistency, avoiding errors and omissions caused by manual operation.

[0159] For example, there are two ways to execute automated tests and generate result data. The first is a standardized serial execution and rule-based data processing method. First, according to the priority order of the test cases, the parameter configuration and execution judgment criteria of each test case are parsed sequentially to generate the corresponding test script. Then, according to the environmental requirements of the test script, dependent components and drivers are installed sequentially, the software and hardware test environment is configured, and environment verification is completed. Next, the target software is driven to execute each test case serially according to the test script, synchronously collecting process data, result data, and exception log data for each test case. After all test cases have been executed, the collected full data is normalized and its integrity is verified according to preset rules. Invalid data is removed by string matching and keyword recognition, and missing fields are filled by default values ​​and association with historical data, finally generating standardized result data. This method uses serial execution and rule-driven logic, with a simple and clear process, high stability, and easy maintenance. It can ensure the repeatability of test results and is suitable for conventional standardized testing scenarios with a small number of test cases.

[0160] The second approach is containerized parallel execution and intelligent data processing. First, all test cases in the test plan are batch-parsed into independent test scripts, each corresponding to an independent test task. Then, based on container technology, an independent container instance is created for each test task, automatically configuring the hardware and software environment within the container, installing dependent components and drivers, and enabling parallel execution of multiple test tasks. Next, a distributed task scheduling engine schedules all test tasks to different computing nodes for parallel execution, monitoring the running status of each container in real time, automatically restarting failed test tasks, and simultaneously collecting process data, result data, and exception log data of all test tasks in parallel through a distributed data acquisition system. After all test tasks have been completed, an anomaly detection algorithm based on machine learning is used to identify invalid data, and missing fields are supplemented through knowledge graph association reasoning. Finally, all data is formatted and normalized to generate standardized result data. This method, employing containerized isolation and distributed parallel computing logic, can fully utilize cluster computing power to improve test execution efficiency, while intelligent data processing technology improves the quality of result data, adapting to complex large-scale test scenarios with a large number of test cases and high computational load.

[0161] Eighth embodiment This embodiment provides an exemplary scheme for multi-dimensional quality assessment and dynamic knowledge base updates. In this example, an assessment dataset is first generated based on the test results of the current software test, combined with user feedback data on search results and test plans. Then, multi-dimensional indicators are calculated and weighted based on the assessment dataset to obtain a suitability score. Next, when the suitability score is lower than a preset threshold, a parameter optimization process is triggered to adjust the search weights and similarity thresholds. Then, effective multimodal data from the current test is extracted to obtain engineering knowledge data. Finally, the engineering knowledge data is embedded into the knowledge base, and the configuration parameters of the engineering semantic graph and corresponding search intent are updated, thereby ultimately achieving a complete closed loop of test quality assessment, adaptive parameter optimization, and dynamic knowledge base updates. Please refer to... Figure 5 , Figure 5 This is a flowchart illustrating the eighth embodiment of the software testing and verification method of this application. Following step S40, steps G11-G15 are also included: Step G11: Based on the test results of this software test, and combined with user feedback data on the search results and the test plan, generate an evaluation dataset.

[0162] Step G12: Based on the evaluation dataset, calculate the quantitative error index, the qualitative reasonableness index, and the matching score of the test scheme in the retrieval results, respectively, and then calculate the fitness score by weighting the quantitative error index, the qualitative reasonableness index, and the matching score.

[0163] Step G13: When the fit score is lower than the preset fit threshold, the parameter optimization process is triggered. A quality loss function is constructed and an adaptive learning rate strategy is used to adjust the retrieval weight and the similarity threshold to complete the iterative optimization of the matching parameters.

[0164] Step G14: Based on the effective multimodal data from this software test, cross-modal feature extraction is performed using the matching parameters to obtain engineering knowledge data.

[0165] Step G15: Embed the engineering knowledge data into the knowledge base, update the node attributes and edge relationships of the engineering semantic graph, and adjust the configuration parameters corresponding to the search intent based on the updated data.

[0166] The evaluation dataset is a structured dataset formed by linking and integrating the results of this software test with user feedback on search results and test plans. It serves as the foundation for quality assessment and parameter optimization. Examples include comparisons between test results and expected results, user relevance ratings of search results, user satisfaction ratings of the test plan, records of user adjustments to the test plan, and user behavior data regarding clicks on search results. Quantitative error indicators quantify the numerical differences between test results and expected results, reflecting the accuracy of the test results. Examples include stress value error, peak temperature error, deformation error, fatigue life error, pressure value error, and modal frequency error. Qualitative rationality indicators evaluate the rationality and effectiveness of test results based on engineering semantic graphs and expert experience rules, reflecting the engineering practicality of the test results. Examples include fault mode recognition accuracy, performance quality assessment accuracy, conformity to physical laws, rationality of boundary conditions, and mesh quality conformity.

[0167] Matching score is an indicator that quantifies the degree of matching between search results and test requirements, as well as the degree of fit between test plans and actual test requirements, reflecting the quality of search and plan generation. Examples include search result relevance score, test plan completeness score, test plan executability score, test plan rationality score, and test plan efficiency score. Adaptability score is a comprehensive quality quantification value obtained by weighted fusion of quantitative error indicators, qualitative rationality indicators, and matching score; it is the core basis for determining whether to trigger the parameter optimization process. The preset adaptation threshold is a pre-set critical value for the adaptation score. When the adaptation score is below this threshold, it indicates that the current matching parameters are not sufficiently adapted and need optimization.

[0168] The quality loss function quantifies the deviation between the current matching parameters and the optimal matching parameters, and is the core computational model for parameter optimization. Examples include mean squared error loss function, cross-entropy loss function, and custom quality loss function. Adaptive learning rate strategies dynamically adjust the learning rate based on the changing trend of the loss function, accelerating convergence while ensuring parameter update stability. Examples include learning rate decay strategies, momentum learning rate strategies, and adaptive moment estimation strategies. Effective multimodal data refers to multimodal data with engineering and reusability value generated during the software testing process, serving as a source of material for knowledge base updates. Examples include new test cases, new simulation results, new fault logs, new expert experience, manually adjusted test plans, and interpretations of new industry standards. Engineering knowledge data is knowledge data obtained by cross-modal feature extraction and structured processing of effective multimodal data, conforming to knowledge base storage specifications, and includes original data, feature vectors, and semantic association information. The configuration parameters for retrieval intent are matching parameters such as retrieval weights and similarity thresholds corresponding to different retrieval intents, and are key parameters for controlling the quality of retrieval matching.

[0169] In this example, the evaluation dataset can be generated using a multimodal data association approach. Alternatively, by introducing a unique test task identifier, all result data and user feedback data from this test can be bound and associated with the test task identifier to automatically generate a structured evaluation dataset, thus completing the construction of the evaluation dataset.

[0170] After constructing the evaluation dataset, the quality assessment process is initiated. Based on the evaluation dataset, quantitative error indicators, qualitative reasonableness indicators, and matching scores between search results and test solutions are calculated. These three indicators are then weighted according to preset weighting coefficients to obtain a suitability score. When the suitability score falls below a preset suitability threshold, a parameter optimization process is triggered. A quality loss function is constructed, and an adaptive learning rate strategy is used to adjust the search weights and similarity thresholds, completing iterative optimization of the matching parameters. Next, effective multimodal data from this software test is extracted, and cross-modal feature extraction is performed using the optimized matching parameters to obtain engineering knowledge data. Finally, the engineering knowledge data is embedded into the knowledge base, updating the node attributes and edge relationships of the engineering semantic graph. Based on the type of updated data, the configuration parameters of the corresponding search intent are adjusted accordingly, thus forming a complete closed loop of quality assessment, parameter optimization, and knowledge base update, continuously improving search matching accuracy and test solution quality.

[0171] For example, there are two ways to achieve multi-dimensional quality assessment and dynamic knowledge base updates. The first is a rule-based static weight closed-loop optimization method. First, according to preset fixed weight coefficients, the weighted values ​​of quantitative error indicators, qualitative reasonableness indicators, and matching scores are calculated and summed to obtain the suitability score. When the suitability score is lower than a preset suitability threshold, a quality loss function based on the difference between the suitability score and the preset threshold is constructed. A gradient descent algorithm with a fixed step size is used to adjust the retrieval weights and similarity thresholds to complete parameter optimization. Next, according to fixed feature extraction rules, cross-modal feature extraction is performed on effective multimodal data to generate engineering knowledge data. Finally, the engineering knowledge data is batch-stored into the knowledge base. The nodes and edges of the engineering semantic graph are supplemented according to preset update rules, and the configuration parameters of the relevant retrieval intent are adjusted according to the type of updated data and preset parameter adjustment rules. This method uses the logic of fixed rules and static weights, has a simple and clear process, high stability, and is easy to implement and maintain. It can meet the closed-loop optimization needs of conventional testing scenarios and is suitable for application scenarios with small data scales and relatively fixed test types.

[0172] The second method is a dynamic weight adaptive closed-loop optimization method based on reinforcement learning. First, the evaluation dataset is input into a pre-trained reinforcement learning agent. The agent dynamically adjusts the weight coefficients of quantitative error indicators, qualitative reasonableness indicators, and matching scores based on the current test scenario and data characteristics, calculating a more suitable fit score. When the fit score is lower than a preset fit threshold, the reinforcement learning agent automatically explores the optimal retrieval weight and similarity threshold adjustment strategy with the goal of maximizing future fit scores. An adaptive learning rate strategy is used to iteratively optimize the parameters, continuously accumulating optimization experience to improve the efficiency and accuracy of subsequent parameter optimization. Next, a cross-modal feature extraction model based on an attention mechanism is used to perform deep feature extraction on effective multimodal data, generating engineering knowledge data with richer semantic information. Finally, the incremental engineering knowledge data is stored in a knowledge base. A graph neural network is used to update the node attributes and edge relationship weights of the engineering semantic graph, and the configuration parameters of the corresponding retrieval intent are automatically adjusted based on the impact of the updated data on the retrieval effect. This method employs reinforcement learning-driven dynamic weights and adaptive optimization logic, which can automatically adjust evaluation criteria and optimization strategies according to different testing scenarios and user needs, continuously improving the overall performance of the system and adapting to application scenarios with large data scales, complex and varied test types, and diverse user needs.

[0173] Furthermore, this embodiment provides a test and verification scheme for retrieval enhancement CAE software based on multiple alternative technical approaches. In the multimodal data vectorization stage, for text data, robustly optimized BERT pretraining models (RoBERTa) and enhanced representation through knowledge integration (ERNIE) can be used to fine-tune CAE-specific corpora to generate high-precision semantic vectors. For image data, the Vision Transformer (ViT) model can be used instead of the Residual Network 50 (ResNet-50) model to extract more globally perceptive visual features. For high-dimensional engineering data, nonlinear dimensionality reduction algorithms such as t-distributed stochastic neighbor embedding (TSNE) and linear discriminant analysis (LDA) can be used instead of principal component analysis (PCA). Component Analysis can better preserve key engineering features. In the retrieval intent classification stage, deep learning models such as Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Bidirectional Encoder Representations from Transformers (BERT) can replace the Improved Support Vector Machine (SVM) multi-classification model, achieving high-accuracy intent classification through training with sufficient CAE domain-specific samples. In the similarity calculation stage, Euclidean distance, Manhattan distance, and other distance calculation methods can replace the cosine similarity algorithm for feature vector similarity calculation. Semantic association similarity can be achieved by using Jaccard coefficients, Pointwise Mutual Information (PMI), and other methods to replace existing calculation methods, enabling effective multi-dimensional similarity evaluation. In the parameter optimization stage, global heuristic optimization algorithms such as genetic algorithms and particle swarm optimization can replace the gradient descent heuristic adjustment strategy, achieving global optimality of matching parameters through population iterative optimization.In the storage architecture stage, a distributed file system (HDFS, Hadoop Distributed File System) can be used to replace the file storage portion of the existing hybrid storage architecture. A graph database (Neo4j, Neo4j Graph Database) can be used to specifically store the engineering semantic graph. This distributed architecture significantly improves the storage and retrieval concurrency processing capabilities of large-scale CAE test data. While the technical alternatives in each of these stages differ in their specific implementation logic, computational efficiency, and applicable scenarios, they all strictly adhere to the core inventive concepts of "Retrieval-Augmented Generation (RAG) method," "CAE domain customization," and "full-process closed-loop optimization." They can equally achieve the core inventive objective of improving the accuracy, efficiency, and intelligence of CAE software testing and verification. Their technical details and implementation methods are all within the protection scope of this invention.

[0174] This application provides a software testing and verification device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the software testing and verification method in Embodiment 1 above.

[0175] The following is for reference. Figure 6 The diagram illustrates a structural schematic of a software testing and verification device suitable for implementing embodiments of this application. The software testing and verification device in these embodiments may include, but is not limited to, mobile terminals such as industrial control computers, high-performance computing servers, and graphics workstations, as well as fixed terminals such as network switches, firewalls, and server racks. Figure 6 The software testing and verification device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0176] like Figure 6As shown, the software testing and verification device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the software testing and verification device. The processing unit 1001, the read-only memory 1002, and the RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the software testing and verification equipment to communicate wirelessly or wiredly with other devices to exchange data. Although a software testing and verification equipment with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0177] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0178] The software testing and verification equipment provided in this application, employing the software testing and verification method described in the above embodiments, can solve the technical problem of low efficiency in computer software testing and verification. Compared with the prior art, the beneficial effects of the software testing and verification equipment provided in this application are the same as those of the software testing and verification method provided in the above embodiments, and other technical features of this software testing and verification equipment are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0179] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0180] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0181] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the software testing and verification method in the above embodiments.

[0182] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, radio frequency (RF), etc., or any suitable combination thereof.

[0183] The aforementioned computer-readable storage medium may be included in the software testing and verification equipment; or it may exist independently and not be assembled into the software testing and verification equipment.

[0184] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a software testing and verification device, the software testing and verification device: responds to software testing instructions, performs feature vectorization processing on multi-dimensional evaluation data to obtain feature vectors, and constructs an engineering semantic graph based on the feature vectors; matches the engineering semantic graph with the core elements of computer-aided engineering testing requirements to determine the retrieval weight and similarity threshold corresponding to this software test; calculates similarity based on the retrieval weight, the similarity threshold, and the engineering semantic graph to determine candidate results, and generates a test plan based on the candidate results and user requirements; and configures the hardware and software testing environment based on the test plan and drives the target software to perform automated testing to obtain the test results of this software test.

[0185] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0186] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation that may be implemented in systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0187] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0188] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., computer programs) for executing the above-described software testing and verification method, thereby solving the technical problem of low efficiency in computer software testing and verification. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the software testing and verification method provided in the above embodiments, and will not be repeated here.

[0189] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A software testing and verification method, characterized in that, The method includes: In response to software testing instructions, feature vectorization is performed on multi-dimensional evaluation data to obtain feature vectors, and an engineering semantic graph is constructed based on the feature vectors. Based on the core elements of computer-aided engineering testing requirements, the engineering semantic graph is matched to determine the retrieval weight and similarity threshold corresponding to this software test; Based on the retrieval weight and the similarity threshold, similarity is calculated using the engineering semantic graph to determine candidate results, and a test plan is generated based on the candidate results and user requirements. Based on the aforementioned test plan, a hardware and software testing environment is configured and the target software is driven to perform automated testing, thereby obtaining the test results of this software test.

2. The software testing and verification method as described in claim 1, characterized in that, The steps of responding to software testing commands, performing feature vectorization on multi-dimensional evaluation data to obtain feature vectors, and constructing an engineering semantic graph based on the feature vectors include: In response to the software testing instructions, multimodal data of the entire process of computer-aided engineering software testing and verification is acquired, and the multimodal data is classified and organized to obtain the multidimensional evaluation data; According to data processing rules and data format, the multidimensional evaluation data is classified and processed to obtain a classification feature set; By using a domain-customized model, the classification feature set is mapped into a unified-dimensional feature vector according to cross-modal feature fusion rules, thus obtaining the feature vector; Based on the correlation between the feature vectors, the relationships between graph nodes and graph edges are determined, and the engineering semantic graph is constructed by combining the graph update mechanism.

3. The software testing and verification method as described in claim 1, characterized in that, The steps of matching the core elements of computer-aided engineering testing requirements with the engineering semantic graph to determine the retrieval weight and similarity threshold corresponding to this software test include: Extract the core elements of the computer-aided engineering testing requirements, and determine the product object, test type, and key parameters for this software test. The product object, test type, and key parameters of this software test are matched with the engineering semantic graph to determine the retrieval intent to which this software test belongs. Based on the search intent matching the preset intent parameter configuration table, the search weight and similarity threshold corresponding to this software test are determined.

4. The software testing and verification method as described in claim 1, characterized in that, The steps of calculating similarity based on the retrieval weight, the similarity threshold, and the engineering semantic graph to determine candidate results, and generating a test plan based on the candidate results and user requirements, include: Based on the retrieval weights and the engineering semantic graph, the feature vector similarity, semantic association similarity, and scene adaptation similarity between the computer-aided engineering testing requirements and the knowledge base data are calculated respectively. The feature vector similarity, semantic association similarity, and scene adaptation similarity are weighted and fused according to a preset similarity weight to obtain the comprehensive similarity of the candidate data; Based on the overall similarity of the candidate data, initial data with an overall similarity higher than the similarity threshold are selected; The initial data is scored based on a comprehensive quality assessment of average similarity, timeliness, and diversity. The initial data are then sorted in descending order according to the comprehensive quality score to obtain the candidate results. The test plan is obtained by analyzing the user requirements and the test elements of the candidate results according to the standard test procedure and adapting the parameters.

5. The software testing and verification method as described in claim 4, characterized in that, Before the step of calculating the feature vector similarity, semantic association similarity, and scene adaptation similarity between the computer-aided engineering test requirements and the knowledge base data based on the retrieval weight and the engineering linguistic graph, the software testing and verification method further includes: Based on a general semantic graph, the core terminology system, parameter constraint rules, and scenario classification standards in the field of computer-aided engineering testing are extracted to form a domain knowledge framework template. The multidimensional evaluation data is classified, labeled, and structured according to the domain knowledge framework template to generate a multimodal data set; Cross-modal feature extraction and fusion processing are performed on a standardized multimodal dataset to obtain a set of domain knowledge feature vectors that correspond to the dimensions of the domain knowledge framework template. The domain knowledge feature vector set is associated and mapped with the domain knowledge framework template, and the node attributes and edge relationships of the engineering semantic graph are supplemented to obtain the knowledge base containing the original data, feature vectors and semantic associations.

6. The software testing and verification method as described in claim 4, characterized in that, The steps of performing a comprehensive quality score on the initial data based on average similarity, timeliness, and diversity, and then sorting the corresponding initial data in descending order according to the comprehensive quality score to obtain the candidate results include: The quality scores of the initial data are calculated according to the three-dimensional scoring rules, including the average similarity, timeliness, and diversity. The quality scores are then correlated with the initial data to obtain the three-dimensional quality scores. Based on the search intent corresponding to this software test, the dimensional weight coefficients of the three-dimensional quality score are adjusted to obtain a quality score weight configuration that adapts to the search intent. The three-dimensional quality scores of the initial data are weighted and fused according to the quality score weight configuration to obtain the comprehensive quality score corresponding to the initial data. The initial data is sorted in descending order of comprehensive quality score, and duplicate or redundant data is removed from the initial data to obtain the candidate results.

7. The software testing and verification method as described in claim 1, characterized in that, The steps of configuring the software and hardware testing environment based on the test plan and driving the target software to perform automated testing to obtain the test results of this software test include: The test cases, parameter configurations, and execution criteria of the test plan are analyzed and converted into test scripts adapted to the target software interface. Configure the hardware and software testing environment according to the environment requirements of the test script, and install the dependent components and drivers to obtain the target testing environment; In the target testing environment, the target software being tested is driven to perform a fully automated test according to the test script, and process data, result data and exception log data are collected during the test. The process data, result data, and exception log data are formatted and their integrity is verified. Invalid data is removed and missing fields are added to generate the result data of this software test.

8. The software testing and verification method as described in claim 1, characterized in that, After the steps of configuring the software and hardware testing environment based on the test plan and driving the target software to perform automated testing to obtain the test results of this software test, the software testing verification method further includes: Based on the test results of this software test, combined with user feedback on the search results and the test plan, an evaluation dataset is generated. Based on the evaluation dataset, the quantitative error index and qualitative reasonableness index of the test results, as well as the matching score of the test scheme in the search results, are calculated respectively. The quantitative error index, the qualitative reasonableness index, and the matching score are then weighted and calculated to obtain the suitability score. When the suitability score is lower than the preset suitability threshold, the parameter optimization process is triggered, a quality loss function is constructed, and an adaptive learning rate strategy is used to adjust the retrieval weight and the similarity threshold to complete the iterative optimization of the matching parameters. Based on the effective multimodal data from this software test, cross-modal feature extraction is performed using the matching parameters to obtain engineering knowledge data. The engineering knowledge data is embedded into the knowledge base, the node attributes and edge relationships of the engineering semantic graph are updated, and the configuration parameters corresponding to the search intent are adjusted in conjunction with the updated data.

9. A software testing and verification device, characterized in that, The software testing and verification device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the software testing and verification method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the software testing and verification method as described in any one of claims 1 to 8.

Images