Automobile software process auditing method and system

Abstract

The invention provides an automobile software process auditing method and system. Relates to the technical field of vehicles, and the method comprises the steps: retrieving first multi-source data generated in each practice in a target process through a first large language model, analyzing a first evidence obtained through retrieval, and determining an achievement condition of an evaluation requirement corresponding to the practice reached by the first evidence; the target process is one of auditing standard setting processes; performing semantic association analysis on second multi-source data of the target process and the association process through a second large language model to generate association data between the two processes; the association process is a process associated with the target process indicated by the auditing standard; and generating an evaluation report of the target process based on the achievement condition of each practice included in the target process and the associated data of the target process and the associated process. Thus, through the two language models, deep semantic understanding and correlation analysis of automobile software process related data are achieved, and limitation of personal errors and configuration rules is effectively reduced.

Description

Technical Field

[0001] This application relates to the field of vehicle technology, and in particular to a method and system for reviewing automotive software processes. Background Technology

[0002] The automotive industry has a set of common audit standards designed to improve the quality standards of the automotive industry. For example, the Automotive Software Process Improvement and Capability Assessment Model (ASPICE) is an authoritative international standard used by the automotive industry to evaluate and improve software development processes. ASPICE defines multiple processes (such as project management, systems engineering, and software engineering) and defines several practices (BP) for each process.

[0003] Furthermore, when reviewing automotive software processes using ASPICE, one approach is manual review. However, manual review suffers from drawbacks such as low efficiency, high cost, strong subjectivity, and poor traceability. Another approach involves hard-coding matching rules, such as `IF (code commit record.associated issue ID = null) THEN violation`, and collecting data for rule matching. If a violation is found, a violation report is generated. However, this approach can only provide binary "yes / no" judgments, lacks semantic understanding capabilities, struggles to handle complex logical data, and has insufficient evaluation depth. Summary of the Invention

[0004] In view of this, this application provides a method and apparatus aimed at improving the ability of semantic understanding and association analysis, overcoming the shortcomings of low efficiency of traditional manual review and lack of semantic understanding ability of rule engines, and improving the level of intelligent review.

[0005] Firstly, this application provides a method for auditing automotive software processes, including:

[0006] Using the first major language model, the first multi-source data generated in each practice of the target process is retrieved, and the first evidence obtained from the retrieval is analyzed to determine the achievement of the evaluation requirements corresponding to the practice; the target process is one of the processes in which the audit criteria are set, and the evaluation requirements are the requirements of the audit criteria for auditing the practice;

[0007] Using a second language model, semantic association analysis is performed on the second multi-source data of the target process and the associated process respectively to generate association data between the target process and the associated process; the associated process is the process that is associated with the audit standard indication and the target process.

[0008] An evaluation report for the target process is generated based on the achievement status of each practice included in the target process and the correlation data between the target process and the related processes.

[0009] Optionally, before retrieving the first multi-source data generated in each practice using the first major language model, analyzing the retrieved first evidence, and determining whether the first evidence meets the evaluation requirements corresponding to the practice, the process includes:

[0010] Input the first training data corresponding to the practice into the first large language model to generate a retrieval expression, which is used to retrieve the first evidence related to the practice;

[0011] The step of retrieving first multi-source data generated in various practices of the target process using a first major language model, and analyzing the first evidence obtained from the retrieval to determine the achievement of the evaluation requirements corresponding to the practice by the first evidence, includes:

[0012] Based on the retrieval formula corresponding to the practice, the first multi-source data is retrieved to obtain the first evidence;

[0013] The first piece of evidence is semantically analyzed using the first major language model to determine the degree of achievement rating and the first judgment reason for achieving the evaluation requirements corresponding to the practice. The first judgment reason is the reason for generating the degree of achievement rating.

[0014] Optionally, the step of performing semantic analysis on the first evidence using a first major language model to determine the degree of achievement rating of the first evidence in fulfilling the evaluation requirements corresponding to the practice includes:

[0015] Based on the evaluation requirements of the audit criteria for the practice, if the first evidence proves that the proportion of the practice that meets the evaluation requirements is greater than a first value, the achievement rating is determined to be F. If the first evidence proves that the proportion of the practice that meets the evaluation requirements is less than or equal to the first value and greater than the second value, the achievement rating is determined to be L. If the first evidence proves that the proportion of the practice that meets the evaluation requirements is less than or equal to the second value and greater than the third value, the achievement rating is determined to be P. If the first evidence proves that the proportion of the practice that meets the evaluation requirements is less than or equal to the third value, the achievement rating is determined to be N.

[0016] Optionally, practices that are rated P or N in terms of achievement during the process of achieving the goal are considered weaknesses, and the weaknesses and their achievement status are recorded in a preliminary weakness list.

[0017] In response to the assessor's confirmation or revision of one or more of the achievement ratings and first judgment reasons for each weakness in the preliminary weakness list, a first label is assigned to the weakness;

[0018] If all weaknesses included in the preliminary weakness list are equipped with a first label, a verified weakness list is obtained and output; and / or, the first language model is iterated based on the achievement of the weaknesses configured with the first label.

[0019] Optionally, the step of performing semantic association analysis on the second multi-source data of the target process and the associated process respectively through a second major language model to generate association data between the target process and the associated process includes:

[0020] Using the second language model, semantic parsing is performed on the first work product of the target process and the second work product of the associated process to determine the second associated item associated with each first associated item. The first work product includes at least one first associated item, and the second associated item is obtained by parsing the data in the second work product based on the data corresponding to the first associated item.

[0021] Generate triples based on the first associated item, the second associated item, and the association type between the first associated item and the second associated item;

[0022] Based on the triples corresponding to each first associated item included in the first working product, an association structure between the first associated item and the second associated item is established; the association structure includes at least one first associated item, at least one second associated item, and an association marker, wherein the association marker is used to indicate that the first associated item and the second associated item corresponding to the association marker are associated.

[0023] Optionally, the association structure is an association matrix; the step of establishing the association structure between the first association item and the second association item based on the triples corresponding to each first association item included in the first working product includes:

[0024] Construct an association matrix, wherein the first row of the association matrix includes at least one first association item, and the first column of the association matrix includes at least one second association item; traverse the triples corresponding to the first association items included in the first work product, determine the second association items associated with each first association item, and set the association marker at the intersection of the column corresponding to the first association item and the row corresponding to the second association item to form the association structure;

[0025] Alternatively, construct an association matrix, wherein the first column of the association matrix includes at least one first association item, and the first row of the association matrix includes at least one second association item; traverse the triples corresponding to the first association items included in the first work product, determine the second association items associated with each first association item, and set the association marker at the intersection of the row corresponding to the first association item and the column corresponding to the second association item to form the association structure.

[0026] Optionally, the method further includes:

[0027] Based on the association structure, if the number of second associations associated with each first association item is less than the number of second associations pre-configured for the first association item, then the missing second association item is determined, and there is a break in the association between the first association item and the missing second association item.

[0028] In the association structure, a break mark is set at the intersection of the first association item and the second association item corresponding to the association break to form a traceability structure.

[0029] Optionally, the step of generating an evaluation report for the target process based on the achievement status of each practice included in the target process and the correlation data between the target process and the related processes includes:

[0030] Practices that are rated P or N in terms of achievement during the process of achieving the stated goal are considered weaknesses, and the weaknesses and their achievement status are recorded in a weakness list; the achievement status includes the achievement rating of each weakness and the first reason for the determination.

[0031] The first associated item corresponding to the break mark in the traceability structure is taken as a weakness, and the weakness and the determination of the weakness are recorded in the weakness list; the determination includes a second determination reason and a second piece of evidence; the second determination reason is used to express the reason for the break of the first associated item and the second associated item corresponding to the break mark, and the second piece of evidence includes the data corresponding to the first associated item in the first work product and the data corresponding to the second associated item in the second work product;

[0032] The evaluation report is generated based on the achievement rating of the practices included in the target process, the list of weaknesses corresponding to the target process, and the traceability structure.

[0033] The method further includes:

[0034] Display the evaluation report on the front-end interface; and / or,

[0035] In response to a click on a fracture marker in the traceability structure on the front-end interface, a second determination reason and second piece of evidence corresponding to the fracture marker are displayed.

[0036] Secondly, this application provides an automotive software process auditing system, comprising:

[0037] The analysis unit is used to retrieve the first multi-source data generated in each practice of the target process using a first major language model, and to analyze the first evidence obtained from the retrieval to determine the achievement of the evaluation requirements corresponding to the practice by the first evidence; the target process is one of the processes in which the audit criteria are set, and the evaluation requirements are the requirements of the audit criteria for auditing the practice;

[0038] The traceability unit is used to perform semantic association analysis on the second multi-source data of the target process and the associated process respectively through the second major language model, and generate association data between the target process and the associated process; the associated process is the process that the audit standard indicates is associated with the target process.

[0039] The generation unit is used to generate an evaluation report of the target process based on the achievement status of each practice included in the target process and the correlation data between the target process and the associated processes.

[0040] Optionally, the system further includes:

[0041] The acquisition and conversion unit is used to acquire the first multi-source heterogeneous data and the second multi-source heterogeneous data, and convert the first multi-source heterogeneous data into first multi-source data understood by the first large language model, and convert the second multi-source heterogeneous data into second multi-source data understood by the second large language model.

[0042] Thirdly, this application provides an apparatus comprising a memory and a processor, the memory for storing instructions or code, and the processor for executing the instructions or code to cause the apparatus to perform an automotive software process auditing method as described in any of the first aspects above.

[0043] Fourthly, this application provides a computer storage medium storing code, wherein when the code is executed, a device running the code implements the automotive software process auditing method described in any of the first aspects above.

[0044] This application provides a method and system for auditing automotive software processes. First, using a first major language model, first multi-source data generated in various practices of a target process is retrieved, and the retrieved first evidence is analyzed to determine whether the first evidence meets the evaluation requirements corresponding to the practice; the target process is one of the processes set by the audit criteria. Then, using a second major language model, semantic association analysis is performed on the second multi-source data of the target process and the associated processes to generate association data between the target process and the associated processes; the associated processes are those processes indicated by the audit criteria and associated with the target process. Finally, based on the achievement status of each practice included in the target process and the association data between the target process and the associated processes, an evaluation report of the target process is generated. Based on the evaluation requirements of the audit criteria for each practice, the first evidence retrieved from the multi-source data is segmented using the first major language model to determine whether the first evidence meets the evaluation requirements of the practice, and to determine whether the first evidence meets the evaluation requirements corresponding to the practice. Furthermore, a second language model is used to perform semantic association analysis on multi-source data of the target process and related processes, mining the correlation data between the target process and the related processes, and then generating an evaluation report by integrating the achievement status of each practice with the correlation data. In this way, by working together with the two language models, the achievement status of the evaluation requirements corresponding to the first piece of evidence and the tracing of the correlation data between the target process and related processes are analyzed, achieving deep semantic understanding and correlation analysis of automotive software process-related data. This improves the intelligence and automation level of the audit, effectively reducing human error and the limitations of configuration rules. Through semantic parsing and process correlation mining of multi-source heterogeneous data, the reliability and traceability of audit standard evaluation are improved, providing data-driven decision support for the continuous optimization of the automotive software development process. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings used in the description of the embodiment or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 A flowchart illustrating an automotive software process review method provided in this application embodiment;

[0047] Figure 2 A flowchart illustrating an optimization of a first major language model provided in an embodiment of this application;

[0048] Figure 3 This is a schematic diagram illustrating the generation process of an association structure provided in an embodiment of this application;

[0049] Figure 4 This is a schematic diagram of the structure of an automotive software process review system provided in an embodiment of this application. Detailed Implementation

[0050] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0051] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0052] Unless otherwise stated, the term "multiple" means two or more. In embodiments of this disclosure, the character " / " indicates that the preceding and following objects are in an "or" relationship. For example, A / B means: A or B. The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or, A and B.

[0053] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0054] The audit standard in this application takes ASPICE as an example. Of course, the method of this application is not only applicable to ASPICE. With appropriate model fine-tuning and data adaptation, it can also be applied to the standard compliance assessment of other industries, such as functional safety ISO 26262 and expected functional safety SOTIF.

[0055] See Figure 1 , Figure 1This application provides a flowchart illustrating a method for reviewing automotive software processes, comprising:

[0056] S101. Using the first major language model, retrieve the first multi-source data generated in each practice of the target process, analyze the first evidence obtained from the retrieval, and determine the achievement of the evaluation requirements corresponding to the practice by the first evidence.

[0057] The target process is one of the processes in which audit standards are set, and the assessment requirements are the requirements for the practice to be audited by the audit standards.

[0058] Optionally, the retrieval is performed by a first language model based on a retrieval expression generated from relevant historical data of the target practice, in order to obtain evidence related to the target practice.

[0059] In one example, the primary language model could be a domain expert model trained on a vast amount of ASPICE official documentation, automotive industry standards, anonymized historical evaluation reports, professional books, and code data.

[0060] The aforementioned first evidence data meets the evaluation requirements of various practices as stipulated by ASPICE, such as whether it constitutes a complete software requirements specification and whether the requirements description is clear and unambiguous.

[0061] In one example, the first major language model parses the review and evaluation requirements for various practices in the automotive software development process from the ASPICE official documentation. It then performs semantic understanding and logical reasoning on the first multi-source data, combined with contextual information, to obtain the first piece of evidence based on the retrieval query corresponding to the event. Furthermore, when understanding the first piece of evidence, the first major language model not only understands explicit content but also infers implicit information, determines whether the first piece of evidence meets the evaluation requirements for the practice, and outputs the achievement status of the practice. For example, this achievement status may include generated structured comments (reasons for judgment) and achievement ratings to support subsequent manual review. Thus, introducing the first major oracle model significantly improves review efficiency and consistency, and reduces the risk of human bias.

[0062] S102. Using the second language model, perform semantic association analysis on the second multi-source data of the target process and the associated process respectively to generate association data between the target process and the associated process; the associated process is the process associated with the target process by the ASPICE instruction.

[0063] For example, ASPICE divides software development into several processes (process areas), such as requirements analysis, software design, software implementation (primarily involving code implementation), and software testing (primarily involving test case testing). Each process is accompanied by specific and implementable practices to ensure the standardization and quality of the entire software development lifecycle. For instance, when the target process is software requirements analysis, the associated processes can be software design, software testing, or other processes related to software requirements.

[0064] The aforementioned associated processes are other processes in ASPICE that have process dependencies or collaborations with the target process. The second major language model performs cross-process semantic understanding and logical association analysis on the second multi-source data involved in the target process and associated processes, generating structured associated data to support full-process traceability verification. For example, in a scenario where the target process is software requirements and the associated process is software design, the model analyzes the relationship between requirement documents and design documents, identifies the mapping relationship between requirement items and design modules, verifies whether each requirement is effectively reflected in the design, and forms structured associated data to trace whether there are any breaks in the association relationship. Furthermore, this associated data can be further used to support change impact analysis; when a requirement is adjusted, the model can quickly locate the affected design modules or test cases.

[0065] S103. Based on the achievement status of each practice included in the target process and the correlation data between the target process and the related processes, generate an evaluation report of the target process.

[0066] Optionally, the assessment report can be presented in a visual format.

[0067] Based on steps S101-S103 above, and according to ASPICE's evaluation requirements for each practice, the first piece of evidence retrieved from multi-source data is segmented using a first major language model to determine whether the first piece of evidence meets the evaluation requirements of the practice, and to determine the achievement status of the evaluation requirements corresponding to the practice. Furthermore, a second major language model is used to perform semantic association analysis on the multi-source data of the target process and related processes, mining the association data between the target process and the related processes, and then combining the achievement status of each practice with the association data to generate an evaluation report. Thus, by working collaboratively with the two major language models, the achievement status of the first piece of evidence in fulfilling the evaluation requirements corresponding to the practice and tracing the association data between the target process and related processes is analyzed, achieving deep semantic understanding and association analysis of automotive software process-related data. This improves the intelligence and automation level of the review process, effectively reducing human error and the limitations of configuration rules. Through semantic parsing of multi-source heterogeneous data and mining of process associations, the reliability and traceability of ASPICE evaluation are improved, providing data-driven decision support for the continuous optimization of the automotive software development process.

[0068] Based on the above embodiments, this application may further include the following before step S101:

[0069] Obtain the first multi-source heterogeneous data; the first multi-source data includes the ASPICE official documentation and the multi-source heterogeneous data corresponding to the target practice; then, parse and vectorize the unstructured data in the first multi-source heterogeneous data (for example, convert the PDF document content into plain text and segment it into blocks, parse the code into an abstract syntax tree), and convert it into a data format that the first large language model can understand, to obtain the first multi-source data.

[0070] The aforementioned acquisition of target practice corresponds to multi-source heterogeneous data, including but not limited to one or more of the following types of data:

[0071] Document Library: Obtain system requirements specifications, software design documents, test plans and reports, meeting minutes, etc. (supports PDF, Word, Excel and other formats).

[0072] Code repository: retrieve source code files, commit logs, and branch management information.

[0073] Project management tools: Acquire user stories, tasks, defects, and change request records.

[0074] Test management tools: retrieve test cases and test execution result records.

[0075] Audio and video data: Acquire interview recordings (converted to text via a speech recognition module).

[0076] In the embodiments of this application, the above Figure 1 There are several possible implementations of step S101, which will be described below. It should be noted that the implementations given below are merely illustrative examples and do not represent all implementations of the embodiments of this application.

[0077] In one example, prior to step S101, the following is included:

[0078] Input the first training data corresponding to the practice into the first large language model to generate a retrieval formula, which is used to retrieve the first evidence related to the practice.

[0079] Optionally, the first training data may include one or more of the following: ASPICE official documentation, automotive industry standards, and historical evaluation reports and code data corresponding to the practice. The first language model understands the descriptions of the practice in the first training data, generates high-precision search expressions, and improves the accuracy of evidence retrieval.

[0080] Furthermore, step S101 above may include:

[0081] Based on the retrieval formula corresponding to the practice, the first multi-source data is retrieved to obtain the first evidence;

[0082] The first piece of evidence is semantically analyzed using the first major language model to determine the degree of achievement rating and the first judgment reason for achieving the evaluation requirements corresponding to the practice. The first judgment reason is the reason for generating the degree of achievement rating.

[0083] The first major language model mentioned above must generate detailed, explainable reasons for each decision and accurately cite supporting evidence.

[0084] In one possible implementation, based on ASPICE's evaluation requirements for the practice, the degree to which the first evidence meets the audit criteria for the practice's evaluation requirements is analyzed, with specific levels as follows:

[0085] Based on the evaluation requirements of the audit criteria for the practice, if the first evidence proves that the proportion of the practice that meets the evaluation requirements is greater than a first value, the achievement rating is determined to be F. If the first evidence proves that the proportion of the practice that meets the evaluation requirements is less than or equal to the first value and greater than the second value, the achievement rating is determined to be L. If the first evidence proves that the proportion of the practice that meets the evaluation requirements is less than or equal to the second value and greater than the third value, the achievement rating is determined to be P. If the first evidence proves that the proportion of the practice that meets the evaluation requirements is less than or equal to the third value, the achievement rating is determined to be N.

[0086] The third value is less than the second value, and the second value is less than the first value. Optionally, the first value can be 85%, the second value can be 50%, and the third value can be 15%.

[0087] In another example, the F-level achievement rating above represents finding sufficient and appropriate primary evidence to prove that the practice was fully implemented, achieving between 85% and 100% of the assessment requirements. The L-level represents finding most of the evidence, but with flaws and deficiencies; most (50% to 85%) of the assessment requirements can be achieved. The P-level may represent that the practice was implemented, but some results were unpredictable. The N-level may represent that no valid evidence was found, or that the primary evidence indicates the practice was not implemented.

[0088] For example, if the achievement rating of "SWE.4 BP4 (test software unit)" is determined to be P (partial achievement), the reason for the determination is that "the unit test for the 'signal filtering module' covers the normal situation, but lacks test cases for boundary values ​​(such as input values ​​that are extremely large or extremely small).

[0089] In one possible implementation, see Figure 2 The diagram shows a flowchart for optimizing the first major language model. Specific steps for optimizing the first major language model may include:

[0090] S201. Practices that are rated P or N in terms of achievement of the stated objectives are considered weaknesses.

[0091] S202. Record the weakness and the status of its achievement in the preliminary weakness list;

[0092] S203. In response to the evaluator's confirmation or revision of one or more of the achievement rating and first judgment reasons in the achievement status of each weakness in the preliminary weakness list, a first label is assigned to the weakness;

[0093] S204. If all the weaknesses included in the preliminary weakness list are equipped with a first label, obtain the verified weakness list and output it.

[0094] S205. Periodically iterate the first language model based on the achievement of the weaknesses configured with the first label.

[0095] Understandably, if revisions are made, the achievement status includes the revised achievement rating and the revised first reason for judgment.

[0096] For example, if the achievement rating of "SWE.4 BP4 (test software unit)" is determined to be P (partially achieved), the reason for the determination is that "the unit test for the 'signal filtering module' covers the normal situation, but lacks test cases for boundary values ​​(such as extremely large or small input values)." Furthermore, the evaluator sees this achievement rating and the corresponding reasoning on the visualization interface and reviews the referenced test cases and code. The evaluator discovers that a test file was missed during the analysis of the first major language model; this file actually contains boundary value tests. The evaluator corrects the achievement rating from P to L or F, and adds to the corresponding first reasoning: "The test file 'unit_test_boundary.c' was missing; this file already contains boundary value tests. Therefore, the practice should be determined as L or F (mostly achieved or fully achieved)." This "manually corrected version" will be used as the more accurate version, labeled with a first tag such as "gold standard," and stored in the training database.

[0097] Of course, if the evaluator determines that the achievement rating and the first judgment reason of the first language model analysis are correct, then the first label such as "gold standard" will be directly marked and stored in the training database.

[0098] Every so often, the system uses all the accumulated "gold standard" data to perform supervised fine-tuning of the existing top-tier language model. After multiple iterations, the model's performance on tasks such as judging the sufficiency of unit tests will increasingly approach or even surpass that of human experts.

[0099] Thus, the first major language model in this application is fine-tuned using professional corpora such as ASPICE and anonymized historical reports to form a dedicated ASPICE expert model. This model possesses the domain knowledge to accurately apply general natural language understanding capabilities to the review of specific industry standards. Furthermore, the achievement rating for each practice can output professionally understandable reasons with specific evidence, which is more conducive to making the judgment process transparent and auditable, and realizing the reliability and adoptability of automated evaluation results.

[0100] In the embodiments of this application, the above Figure 1 There are several possible implementations of step S102, which will be described below. It should be noted that the implementations given below are merely illustrative examples and do not represent all implementations of the embodiments of this application.

[0101] See Figure 3 The diagram illustrates a process for generating an association structure. Step S102 may include:

[0102] S300: Obtain the first work product of the target project and the second work product corresponding to the associated process, and convert the data of the first work product and the second work product respectively to obtain a data format that can be understood by the second language model;

[0103] For example, the work product corresponding to the requirements process can be a system requirements specification; the work product corresponding to the software design process can be a software design document; the work product corresponding to the software implementation process can be source code files; and the work product corresponding to the software testing process can be test cases.

[0104] Of course, the original format of the working product in this application can be multi-source heterogeneous data, such as documents (supporting PDF, Word, Excel and other formats), code, audio, video and other data. Then, this application will convert the data of various formats into a format that the second large language model can understand, such as converting the content of PDF documents into plain text and splitting it into blocks, and parsing the code into an abstract syntax tree, thereby preparing for subsequent large language model analysis.

[0105] S301. Using the second language model, perform semantic parsing on the first work product of the target process and the second work product of the associated process to determine the second associated item associated with each first associated item. The first work product includes at least one first associated item, and the second associated item is obtained from the second work product based on the data corresponding to the first associated item.

[0106] For example, the second largest language model is used to analyze multiple first related items (multiple requirements) in the first work product (such as the system requirements specification) of the target process, and based on their semantic content, second related items (software modules) that logically correspond to each requirement are found in the second work product (such as the software design document) of the related process.

[0107] For example, the second language model analyzes a requirement in the "System Requirements Specification" that "the system should respond to the accelerator pedal signal within 100ms" and a design module "signal processing timing control module" in the "Software Design Document". Then, the model generates a judgment, such as: "Requirement: The system should respond to the accelerator pedal signal within 100ms (Requirement ID: SYS_REQ_001) is implemented by 'signal processing timing control module' (Design ID: DES_MOD_005)". Of course, if there is no corresponding ID number, it can also be represented based on the relevant description in the source document.

[0108] S302. Generate a triplet based on the first associated item, the second associated item, and the association type between the first associated item and the second associated item.

[0109] For example, a triplet structure can be <first associated item, relation type, second associated item>, such as:<SYS_REQ_001, implemented_by, DES_MOD_005> .

[0110] S303. Based on the triples corresponding to each first associated item included in the first working product, establish an association structure between the first associated item and the second associated item; the association structure includes at least one first associated item, at least one second associated item, and an association marker, wherein the association marker is used to indicate that the first associated item and the second associated item corresponding to the association marker are associated.

[0111] The triples analyzed from the first and second working products can be stored in a database, and then the data can be structurally assembled to form an association structure, and visualization processing can be selected.

[0112] For example, the above association structure can be an association matrix or an association two-dimensional table, etc.

[0113] In one example, the above association structure is an association matrix, and step S303 can be:

[0114] Construct an association matrix, wherein the first row of the association matrix includes at least one first association item, and the first column of the association matrix includes at least one second association item; traverse the triples corresponding to the first association items included in the first work product, determine the second association items associated with each first association item, and set the association marker at the intersection of the column corresponding to the first association item and the row corresponding to the second association item to form the association structure.

[0115] In another example, step S303 above can be:

[0116] Construct an association matrix, wherein the first column of the association matrix includes at least one first association item, and the first row of the association matrix includes at least one second association item; traverse the triples corresponding to the first association items included in the first work product, determine the second association items associated with each first association item, and set the association marker at the intersection of the row corresponding to the first association item and the column corresponding to the second association item to form the association structure.

[0117] In another example, step S303 above can dynamically define rows and columns based on the range of data parsing.

[0118] The aforementioned association matrix explicitly expresses the semantic relationship between the first and second association items through the association markers at the intersections of rows and columns, thereby supporting subsequent automated traceability analysis and consistency verification.

[0119] S304. Based on the association structure, if the number of second association items associated with each first association item is less than the number of second association items pre-configured for the first association item, then the missing second association item is determined, and there is a break in the association between the first association item and the missing second association item.

[0120] The second associated item pre-configured for the first associated item can be a first associated item and a second associated item that have been verified based on human experience or historical data, used to measure the actual parsed association structure. When an association break is detected, the associated item corresponding to the break can be identified, which facilitates subsequent manual review and ensures the synergy and integrity between work products.

[0121] S305. In the association structure, a break mark is set at the position where the first association item and the second association item corresponding to the association break intersect, forming a traceability structure.

[0122] Based on steps S301-S305 above, this application constructs a traceability chain between processes for ASPICE auditing. This method understands the semantic logic between work products through a large language model. For example, it can sequentially analyze each of the four processes—requirements, software design, software code implementation, and software testing—in pairs, automatically and dynamically establishing and verifying the bidirectional traceability relationship between each process. For instance, it can detect when a requirement is not covered by any design module (forward traceability break) or when a test case cannot be traced back to any requirement (backward traceability break), thus solving the problem that existing technologies cannot automatically automate integrity and consistency checks.

[0123] In one example, a traceability structure is established between the requirements and design processes (using a two-dimensional table as an example of the traceability structure):

[0124] First, the second language model analyzes the work products corresponding to each of the two processes.

[0125] Optionally, the product of this work can be heterogeneous data. For example, in parsing the System Requirements Specification (text), there is a requirement: "REQ_ID: SYS-101: The vehicle should be able to adjust the steering assist gain according to the driving mode (economy, sport) selected by the driver." In parsing the Software Architecture Design document, there is a passage: "'Driving Mode Processing Module' receives the mode selection signal and calculates the corresponding 'desired assist gain value'." This document also includes an architecture diagram showing a steering assist control module named "SteeringAssistCtrl" with an arrow pointing to the "PowerManagement" power management and distribution module.

[0126] Then, the second language model performs correlation analysis to determine whether the design meets the requirement SYS-101.

[0127] Specifically, regarding text-to-text association: the model understands that the core of requirement SYS-101 is that "driving mode" affects "steering assist". The model found the keywords "driving mode processing module" and "expected assist gain value" in the design document text, initially determining that the two are related. Furthermore, regarding visual and textual association: the model further analyzes the architecture diagram and identifies "SteeringAssistCtrl" (steering assist control module). It then connects this module in the diagram with the "expected assist gain value" in the document text, inferring that the output of the "driving mode processing module" is ultimately passed to "SteeringAssistCtrl" responsible for implementing steering assist. Finally, generating the association conclusion: based on the above analysis, the model generates a structured triplet output: <First association item: SYS-101, Relationship type: implemented_by, Second association item: SteeringAssistCtrl (from architecture diagram) / Driving mode processing module (from text)>, with the second reasoning: the function in requirement SYS-101 regarding adjusting steering assist based on driving mode is decided by the "driving mode processing module" in the design, and its output expected gain value is passed to the "SteeringAssistCtrl" module for implementation. The second piece of evidence comes from the text description on page X and the architecture diagram on page Y of the design document.

[0128] Furthermore, generate the associated structure, for example, create the traceability structure shown in Table 1:

[0129] Table 1

[0130]

[0131] SYS-101 has established associations with both design modules (driving mode processing module and SteeringAssistCtrl), and association markers such as "√" can be added to demonstrate the sufficiency of the design.

[0132] Furthermore, in one example, as shown in Table 2 above, if SYS-102 and SYS-101 are the same requirement, and if the association structure of SYS-101 is verified to be correct, in the case where the association between SYS-102 and the above two design modules is broken, a break mark can be marked, such as "×", and SYS-102 can be regarded as a weakness in the requirement process.

[0133] In another example, as shown in Table 2 below, if the second associated item (design module) associated with multiple first associated items (requirement SYS-101 and SYS-101) in the table is consistent, then the intersection of the corresponding row and column can be kept empty when the association breaks. For example, when the association between SYS-102 and the above two design modules breaks, the corresponding table is empty, then SYS-102 is a weak item in the requirement process.

[0134] Table 2

[0135]

[0136] Finally, SYS-102 can be added to the weakness list to draw the evaluator's attention. Correspondingly, a second reason for judgment can be configured: the function in requirement SYS-102 regarding adjusting steering assist based on driving mode was not decided by the "Driving Mode Processing Module" in the design, nor was the expected output gain value passed to the "SteeringAssistCtrl" module for implementation. The second piece of evidence is the lack of relevant documentation in the design documents.

[0137] Based on the above examples, it is understood that the introduction of the first and second major language models in this application can realize cross-modal association analysis. For example, the data sources and uses of multimodal data can be seen in Table 3 below:

[0138] Table 3

[0139]

[0140] Furthermore, when this application uses the first or second largest language model to understand the corresponding data, for example, it can associate visual information with text to confirm whether the textual description in the design document is consistent with the architecture diagram; it can associate code with text to verify whether a piece of code correctly implements a requirement described in natural language; and it can associate voice with data: through interview recordings, it can confirm whether the root cause and analysis process of a problem marked as "resolved" in the issue tracking system were fully discussed in the meeting.

[0141] Thus, this application enhances the ability to perform multimodal fusion analysis of data by introducing a large language model, which is completely unattainable by existing rule engine-based technologies, enabling the review process to penetrate into the substantive level.

[0142] Based on the above embodiments, step S103 may include:

[0143] First, practices that are rated P or N in terms of achievement during the process of achieving the goal are considered weaknesses, and the weaknesses and their achievement status are recorded in a weakness list; the achievement status includes the achievement rating of each weakness and the first reason for the determination.

[0144] Then, the first associated item corresponding to the break mark in the traceability structure is taken as a weakness, and the weakness and the determination of the weakness are recorded in the weakness list; the determination includes a second determination reason and a second piece of evidence; the second determination reason is used to express the reason for the break of the first associated item and the second associated item corresponding to the break mark, and the second piece of evidence includes the data corresponding to the first associated item in the first work product and the data corresponding to the second associated item in the second work product;

[0145] Finally, based on the achievement rating of the practices included in the target process, the list of weaknesses corresponding to the target process, and the traceability structure, the evaluation report is generated.

[0146] Furthermore, after step S103 above, the following may also be included:

[0147] Display the evaluation report on the front-end interface; and / or,

[0148] In response to a click on a fracture marker in the traceability structure on the front-end interface, a second determination reason and second piece of evidence corresponding to the fracture marker are displayed.

[0149] For example, when displaying Table 2 above, clicking the break mark × corresponding to SYS-102 and the driving mode processing module reveals the second reason for judgment: the function in requirement SYS-102 regarding adjusting steering assist according to the driving mode was not decided by the "driving mode processing module" in the design. The second piece of evidence is that the design documents do not contain relevant records.

[0150] Based on the above embodiments, the advantages of this application may include:

[0151] (1) Improve efficiency: The evidence analysis and judgment are realized by using a large language model, instead of human analysis and review, which reduces labor costs and improves processing efficiency.

[0152] (2) Objectivity and consistency: Since the entire review process is carried out by the same trained model, the subjective judgment bias between different human evaluators can be eliminated, thereby ensuring the high consistency, objectivity and reproducibility of the evaluation results.

[0153] (3) Wider coverage: It can comprehensively process relevant data, rather than limited sampling, thus it is better able to discover deep-seated, systemic problems and process trends that are easily overlooked by manual review and hidden in a large amount of data. It solves the fundamental defect of existing rule matching technology that cannot understand and reason about the content of unstructured documents, and realizes the leap from keyword matching to semantic conformity judgment, thereby providing a more comprehensive evaluation perspective.

[0154] (4) Real-time process monitoring and immediate feedback: It is conducive to integration into the development pipeline, realizing continuous compliance checks in the development process, issuing early warnings at the initial stage of problems, thereby shifting quality assurance from post-audit to in-process protection.

[0155] (5) Lowering the assessment threshold: The large language model absorbs and understands a large amount of data, which is conducive to forming a high-quality audit model. It enables ordinary engineers or assistant assessors to complete high-quality process assessments with the assistance of the system, thereby accelerating the large-scale replication and inheritance of ASPICE's valuable knowledge and experience.

[0156] The above describes some specific implementations of the automotive software process review method provided in this application. Based on this, this application also provides a corresponding apparatus. The apparatus provided in this application will be described below from the perspective of functional modularity.

[0157] See Figure 4 The diagram shows the structure of an automotive software process audit system. The automotive software process audit system includes:

[0158] Analysis unit 401 is used to retrieve first multi-source data generated in each practice of the target process through a first large language model, and to analyze the first evidence obtained by the retrieval to determine the achievement of the evaluation requirements corresponding to the practice by the first evidence; the target process is one of the processes in which the audit criteria are set.

[0159] The traceability unit 402 is used to perform semantic association analysis on the second multi-source data of the target process and the associated process respectively through the second language model, and generate association data between the target process and the associated process; the associated process is the process that the audit standard indicates is associated with the target process.

[0160] The generation unit 403 is used to generate an evaluation report of the target process based on the achievement status of each practice included in the target process and the correlation data between the target process and the associated processes.

[0161] Based on the aforementioned system, the two language models in the analysis unit 401 and the tracing unit 402 of this application work collaboratively. The analysis unit 401 analyzes the achievement of the evaluation requirements corresponding to the practice based on the first piece of evidence, and the tracing unit 402 traces the associated data of the target process and related processes. This achieves deep semantic understanding and correlation analysis of automotive software process-related data, improving the intelligence and automation level of the audit and effectively reducing human error and the limitations of configuration rules. Thus, through semantic parsing of multi-source data and mining of process correlations, the reliability and traceability of audit standard evaluation are improved, providing data-driven decision support for the continuous optimization of the automotive software development process.

[0162] In one possible implementation, the above system further includes:

[0163] The acquisition and conversion unit is used to acquire the first multi-source heterogeneous data and the second multi-source heterogeneous data, and convert the first multi-source heterogeneous data into first multi-source data understood by the first large language model, and convert the second multi-source heterogeneous data into second multi-source data understood by the second large language model.

[0164] It is understandable that the aforementioned second multi-source heterogeneous data is the first working product of the target project D and the second working product corresponding to the association process obtained in step S300 above.

[0165] Optionally, the acquisition and conversion unit may include multiple interfaces for receiving and acquiring the first multi-source heterogeneous data and the second multi-source heterogeneous data.

[0166] In one possible implementation, the analysis unit 401 is further configured to input the first training data corresponding to the practice into the first large language model and generate a retrieval expression, which is used to retrieve the first evidence related to the practice.

[0167] Furthermore, the analysis unit 401 is specifically used to retrieve the first multi-source data based on the retrieval formula corresponding to the practice to obtain the first evidence; and to perform semantic analysis on the first evidence through the first large language model to determine the achievement rating and the first judgment reason of the first evidence in achieving the evaluation requirements corresponding to the practice, wherein the first judgment reason is the reason for generating the achievement rating.

[0168] In one possible implementation, the analysis unit 401 is specifically used to assess the practice based on the audit criteria. If the first evidence proves that the proportion of the practice meeting the assessment requirements is greater than a first value, the achievement rating is determined to be F. If the first evidence proves that the proportion of the practice meeting the assessment requirements is less than or equal to the first value but greater than a second value, the achievement rating is determined to be L. If the first evidence proves that the proportion of the practice meeting the assessment requirements is less than or equal to the second value but greater than a third value, the achievement rating is determined to be P. If the first evidence proves that the proportion of the practice meeting the assessment requirements is less than or equal to the third value, the achievement rating is determined to be N.

[0169] In one possible implementation, the analysis unit 401 is specifically configured to consider practices with an achievement rating of P or N in the target process as weaknesses, and record the weaknesses and their achievement status in a preliminary weakness list; in response to the evaluator's confirmation or revision of one or more of the achievement ratings and first judgment reasons for each weakness in the preliminary weakness list, configure a first label for the weakness; if all weaknesses included in the preliminary weakness list are equipped with first labels, obtain and output a verified weakness list; and / or, perform feedback iteration on the first language large model based on the achievement status of the weaknesses configured with first labels.

[0170] In one possible implementation, the traceability unit includes:

[0171] The association analysis subunit is used to perform semantic parsing on the first work product of the target process and the second work product of the association process through the second language model, and to determine the second association item associated with each first association item. The first work product includes at least one first association item, and the second association item is obtained by parsing the data in the second work product based on the data corresponding to the first association item.

[0172] The triplet subunit is used to generate triplets based on the first associated item, the second associated item, and the association type between the first associated item and the second associated item;

[0173] An assembly subunit is used to establish an association structure between the first association item and the second association item based on the triplet corresponding to each first association item included in the first working product; the association structure includes at least one first association item, at least one second association item, and an association marker, wherein the association marker is used to indicate that the first association item and the second association item corresponding to the association marker are associated.

[0174] The assembly subunit is used to construct an association matrix, wherein the first row of the association matrix includes at least one first association item, and the first column of the association matrix includes at least one second association item; iterates through the triples corresponding to the first association items included in the first work product, determines the second association items associated with each first association item, and sets the association marker at the intersection of the column corresponding to the first association item and the row corresponding to the second association item to form the association structure; or, it constructs an association matrix, wherein the first column of the association matrix includes at least one first association item, and the first row of the association matrix includes at least one second association item; iterates through the triples corresponding to the first association items included in the first work product, determines the second association items associated with each first association item, and sets the association marker at the intersection of the row corresponding to the first association item and the column corresponding to the second association item to form the association structure.

[0175] The assembly subunit is further configured to, based on the association structure, determine the missing second association if the number of second associations associated with each first association is less than the number of second associations pre-configured for the first association, and an association break exists between the first association and the missing second association; in the association structure, a break mark is set at the position where the first association and the second association corresponding to the association break intersect, forming a traceability structure.

[0176] In one possible implementation, the generation unit is specifically configured to treat practices with an achievement rating of P or N in the target process as weaknesses, and record the weaknesses and their achievement status in a weakness list; the achievement status includes the achievement rating of each weakness and a first reason for determination; treat the first associated item corresponding to the break marker in the traceability structure as a weakness, and record the weakness and its determination status in a weakness list; the determination status includes a second reason for determination and second evidence; the second reason for determination is used to indicate the reason for the break in the association between the first associated item and the second associated item corresponding to the break marker, and the second evidence includes the data corresponding to the first associated item in the first work product and the data corresponding to the second associated item in the second work product; and generate the evaluation report based on the achievement rating of the practices included in the target process, the weakness list corresponding to the target process, and the traceability structure.

[0177] In one possible implementation, the system further includes a visualization interface for displaying the evaluation report on a front-end interface; and / or, in response to a click on a fracture marker in the traceability structure on the front-end interface, displaying a second determination reason and second evidence corresponding to the fracture marker.

[0178] This application also provides corresponding devices and computer storage media for implementing the solutions provided in this application.

[0179] The device includes a memory and a processor. The memory stores instructions or code, and the processor executes the instructions or code to enable the device to perform an automotive software process auditing method according to any embodiment of this application.

[0180] The computer storage medium stores code, and when the code is run, the device running the code implements an automotive software process auditing method according to any embodiment of this application.

[0181] In the embodiments of this application, the terms "first" and "second" (if they exist) are used only as name identifiers and do not represent the order of first and second.

[0182] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps in the methods of the above embodiments can be implemented by means of software plus a general-purpose hardware platform. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as a read-only memory (ROM) / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, a server, or a network communication device such as a router) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0183] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0184] The above description is merely an exemplary implementation of this application and is not intended to limit the scope of protection of this application.

Claims

1. A method for reviewing automotive software processes, characterized in that, include: Using the first major language model, the first multi-source data generated in each practice of the target process is retrieved, and the first evidence obtained from the retrieval is analyzed to determine the achievement of the evaluation requirements corresponding to the practice; the target process is one of the processes in which the audit criteria are set, and the evaluation requirements are the requirements of the audit criteria for auditing the practice; Using a second language model, semantic association analysis is performed on the second multi-source data of the target process and the associated process respectively to generate association data between the target process and the associated process; the associated process is the process that is associated with the audit standard indication and the target process. An evaluation report for the target process is generated based on the achievement status of each practice included in the target process and the correlation data between the target process and the related processes.

2. The method according to claim 1, characterized in that, Before retrieving the first multi-source data generated in each practice using the first major language model, analyzing the first evidence obtained from the retrieval, and determining the achievement of the evaluation requirements corresponding to the practice by the first evidence, the process includes: Input the first training data corresponding to the practice into the first large language model to generate a retrieval expression, which is used to retrieve the first evidence related to the practice; The step of retrieving first multi-source data generated in various practices of the target process using a first major language model, and analyzing the first evidence obtained from the retrieval to determine the achievement of the evaluation requirements corresponding to the practice by the first evidence, includes: Based on the retrieval formula corresponding to the practice, the first multi-source data is retrieved to obtain the first evidence; The first piece of evidence is semantically analyzed using the first major language model to determine the degree of achievement rating and the first judgment reason for achieving the evaluation requirements corresponding to the practice. The first judgment reason is the reason for generating the degree of achievement rating.

3. The method according to claim 2, characterized in that, The step of performing semantic analysis on the first evidence using a first major language model to determine the degree of achievement rating of the first evidence in fulfilling the evaluation requirements corresponding to the practice includes: Based on the evaluation requirements of the audit criteria for the practice, if the first evidence proves that the proportion of the practice that meets the evaluation requirements is greater than a first value, the achievement rating is determined to be F. If the first evidence proves that the proportion of the practice that meets the evaluation requirements is less than or equal to the first value and greater than the second value, the achievement rating is determined to be L. If the first evidence proves that the proportion of the practice that meets the evaluation requirements is less than or equal to the second value and greater than the third value, the achievement rating is determined to be P. If the first evidence proves that the proportion of the practice that meets the evaluation requirements is less than or equal to the third value, the achievement rating is determined to be N.

4. The method according to claim 3, characterized in that, Practices that are rated P or N in terms of achievement of the stated objectives are considered weaknesses, and the weaknesses and their achievement status are recorded in a preliminary weakness list. In response to the assessor's confirmation or revision of one or more of the achievement ratings and first judgment reasons for each weakness in the preliminary weakness list, a first label is assigned to the weakness; If all weaknesses included in the preliminary weakness list are equipped with a first label, a verified weakness list is obtained and output; and / or, the first language model is iterated based on the achievement of the weaknesses configured with the first label.

5. The method according to claim 3, characterized in that, The step involves performing semantic association analysis on the second multi-source data of the target process and the associated process using a second major language model to generate association data between the target process and the associated process, including: Using the second language model, semantic parsing is performed on the first work product of the target process and the second work product of the associated process to determine the second associated item associated with each first associated item. The first work product includes at least one first associated item, and the second associated item is obtained by parsing the data in the second work product based on the data corresponding to the first associated item. Generate triples based on the first associated item, the second associated item, and the association type between the first associated item and the second associated item; Based on the triples corresponding to each first associated item included in the first working product, an association structure between the first associated item and the second associated item is established; the association structure includes at least one first associated item, at least one second associated item, and an association marker, wherein the association marker is used to indicate that the first associated item and the second associated item corresponding to the association marker are associated.

6. The method according to claim 5, characterized in that, The association structure is an association matrix; the step of establishing the association structure between the first association item and the second association item based on the triples corresponding to each first association item included in the first working product includes: Construct an association matrix, wherein the first row of the association matrix includes at least one first association item, and the first column of the association matrix includes at least one second association item; traverse the triples corresponding to the first association items included in the first work product, determine the second association items associated with each first association item, and set the association marker at the intersection of the column corresponding to the first association item and the row corresponding to the second association item to form the association structure; Alternatively, construct an association matrix, wherein the first column of the association matrix includes at least one first association item, and the first row of the association matrix includes at least one second association item; traverse the triples corresponding to the first association items included in the first work product, determine the second association items associated with each first association item, and set the association marker at the intersection of the row corresponding to the first association item and the column corresponding to the second association item to form the association structure.

7. The method according to claim 5 or 6, characterized in that, The method further includes: Based on the association structure, if the number of second associations associated with each first association item is less than the number of second associations pre-configured for the first association item, then the missing second association item is determined, and there is a break in the association between the first association item and the missing second association item. In the association structure, a break mark is set at the intersection of the first association item and the second association item corresponding to the association break to form a traceability structure.

8. The method according to claim 7, characterized in that, The evaluation report for the target process is generated based on the achievement status of each practice included in the target process and the correlation data between the target process and the related processes, including: Practices that are rated P or N in terms of achievement during the process of achieving the stated goal are considered weaknesses, and the weaknesses and their achievement status are recorded in a weakness list; the achievement status includes the achievement rating of each weakness and the first reason for the determination. The first associated item corresponding to the break mark in the traceability structure is taken as a weakness, and the weakness and the determination of the weakness are recorded in the weakness list; the determination includes a second determination reason and a second piece of evidence; the second determination reason is used to express the reason for the break of the first associated item and the second associated item corresponding to the break mark, and the second piece of evidence includes the data corresponding to the first associated item in the first work product and the data corresponding to the second associated item in the second work product; The evaluation report is generated based on the achievement rating of the practices included in the target process, the list of weaknesses corresponding to the target process, and the traceability structure. The method further includes: Display the evaluation report on the front-end interface; and / or, In response to a click on a fracture marker in the traceability structure on the front-end interface, a second determination reason and second piece of evidence corresponding to the fracture marker are displayed.

9. A software process auditing system for automobiles, characterized in that, include: The analysis unit is used to retrieve the first multi-source data generated in each practice of the target process using a first major language model, and to analyze the first evidence obtained from the retrieval to determine the achievement of the evaluation requirements corresponding to the practice by the first evidence; the target process is one of the processes in which the audit criteria are set, and the evaluation requirements are the requirements of the audit criteria for auditing the practice; The tracing unit is used to perform semantic association analysis on the second multi-source data of the target process and the associated process respectively through the second language model, and generate association data between the target process and the associated process. The associated process is the process by which the audit criteria indicate the connection to the target process; The generation unit is used to generate an evaluation report of the target process based on the achievement status of each practice included in the target process and the correlation data between the target process and the associated processes.

10. The system according to claim 9, characterized in that, The system also includes: The acquisition and conversion unit is used to acquire the first multi-source heterogeneous data and the second multi-source heterogeneous data, and convert the first multi-source heterogeneous data into first multi-source data understood by the first large language model, and convert the second multi-source heterogeneous data into second multi-source data understood by the second large language model.

Images