An AI large model-based software development application data processing method

Abstract

The application belongs to the field of software development application data processing, and particularly relates to a software development application data processing method based on an AI large model, which comprises the following steps: adopting a unified collection and text standardization processing method for first-line feedback and work order data, cleaning and labeling description information from each terminal; performing semantic analysis through an AI large model to extract elements such as equipment, process, function and problem type; adopting a hierarchical modeling and relationship extraction method for equipment, process and function information to construct an equipment-process-function relationship network; performing path traversal and weighted aggregation on text associated nodes to calculate the influence score of each software function; performing task template filling and historical case retrieval on high-score functions; and performing effect evaluation and parameter updating on the running results of the online development task; the application can uniformly process multi-source application data, quantify the influence of software functions and automatically generate development tasks, and realizes fine research and development decision-making.

Description

Technical Field

[0001] This invention belongs to the field of software development application data processing, specifically relating to a software development application data processing method based on a large AI model. Background Technology

[0002] The Chinese patent application (CN202510560211.9) discloses a method for processing software development application data based on an AI large-scale model. This method deploys an anomaly detection model at the data access layer to dynamically detect and clean software development application data in real time; optimizes the scoring model based on historical security event feedback and classifies the software development application data; performs anti-correlation desensitization training on the software development application data and applies homomorphic encryption; periodically conducts correlation attack simulation tests and adjusts parameters; manages keys in a distributed manner and performs auditing and tracing based on blockchain; accesses a regulatory database based on a dynamic compliance engine, parses regulations to generate a rule base, deploys checkpoints at key data points for real-time detection and blocking, and performs secure erasure verification; combines large-scale model prediction to dynamically schedule resources and monitors costs in real time through a monitoring system; deploys nodes on multiple platforms, encrypts and aggregates parameters to update the model, and periodically scans for vulnerabilities.

[0003] While this method can improve data preprocessing efficiency and strengthen sensitive data protection to some extent in the field of software development and application data processing, in industrial manufacturing enterprises, the large amount of text data generated by multiple systems such as equipment monitoring, work order management, quality traceability, and energy consumption management is still scattered and stored in MES, ERP, operation and maintenance systems, and demand management platforms. Frontline feedback often remains in chat tools or scattered documents, lacking unified archiving and structured descriptions, making it difficult for R&D personnel to grasp the background and scope of the problem in a timely and accurate manner. Existing solutions usually only perform statistics within a single system and cannot follow the "equipment →" path. The "process → software function" link makes it difficult to assess the actual impact of software functions on the production scenario when tracing the problem propagation path. Manually extracting problems and prioritizing them based on text feedback is time-consuming and prone to overlooking related equipment and process information, making it difficult to accumulate reusable knowledge. At the same time, tagging methods based on keyword matching or traditional machine learning lack a unified relationship network across "equipment-process-software function", have limited semantic understanding of colloquial, mixed Chinese and English and abbreviations in the field, make it difficult to quantify the impact of text events on software functions, and also make it difficult to automatically generate structured development tasks and form a self-learning closed loop with the online results.

[0004] To address the aforementioned issues, this invention proposes a data processing method for software development applications based on an AI large-scale model. This method involves unified collection and standardization of multi-source text data, followed by semantic extraction of equipment, processes, software functions, and problem types using an AI large-scale model. Based on this, an equipment-process-function relationship network is constructed, and a propagation path influence scoring algorithm is employed to quantify the impact of each software function. A prioritized set of structured development tasks is generated based on this score, and parameters are updated in conjunction with the software launch results, forming a self-learning data processing mechanism. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a software development application data processing method based on AI large model, aiming to solve the problems of scattered multi-source application data, difficulty in quantifying the impact of software functions, and automatic generation of development tasks in industrial manufacturing enterprises.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] A data processing method for software development applications based on AI large-scale models, comprising:

[0009] Step S1: Collect frontline feedback and work order data in a unified manner and perform text standardization processing to generate a comprehensive data set;

[0010] Step S2: Based on the comprehensive data set, semantic parsing is performed using an AI large model to extract elements related to business objects and problem types, and generate a standardized text set with labels.

[0011] Step S3: Based on a standardized text set, a device-process-function relationship network is generated by performing hierarchical modeling and relationship extraction on business object information.

[0012] Step S4: Based on the functional relationship network, a score set is generated by traversing and weighting the text association nodes and calculating the impact score of each software function.

[0013] Step S5: Based on the rating set, generate a development task set by performing task template population and historical case retrieval on the rating set;

[0014] Step S6 involves evaluating the performance and updating the parameters of the development task after it goes live, thus forming a self-learning closed-loop model.

[0015] As a preferred implementation, the specific steps for uniformly collecting frontline feedback and work order data, performing text standardization processing, and generating a comprehensive dataset are as follows:

[0016] By configuring data collection interfaces between chat tools, email systems, work order systems, and data platforms, and retrieving field data containing problem descriptions, equipment numbers, process steps, occurrence times, and responsible persons at preset time intervals, the collected text data is encoded in a unified format, removing emojis and meaningless whitespace. Long sentences are broken down and labeled with data sources and timestamps line by line. Following a unified data model, the text is mapped to standard fields for problem descriptions, scenario information, object identifiers, and attachment references. Missing fields are filled with placeholders, and the original data primary keys are recorded. The processed multi-source data is written into the same dataset, forming a comprehensive data set that can be used for subsequent semantic analysis.

[0017] As a preferred implementation, the specific steps for using a large AI model to perform semantic parsing, extract elements related to business objects and problem types, and generate a standardized text set with labels are as follows:

[0018] By constructing a semantic extraction template for industrial manufacturing scenarios, the elements to be identified are defined as structured slots representing equipment identifiers, process steps, software functions, problem types, scope of impact, and occurrence scenarios. Example expressions and mandatory rules are configured for each slot. Each piece of text in the comprehensive dataset is input into a pre-trained AI model in the form of "contextual description + extraction instructions," and the output is parsed to obtain candidate equipment names, process names, function names, and problem type labels. Based on pre-maintained equipment ledgers, process dictionaries, and function lists, similarity matching and alias mapping are performed on candidate names to uniformly map colloquial abbreviations and spelling errors to standard codes. For missing or conflicting slots, rules are used to complete them according to timestamps, responsible persons, and affiliated systems, and confidence level markers are recorded. Elements extracted and validated by the AI ​​model are written into corresponding fields and stored together with the original text in a standardized text record, forming a standardized text set with semantic tags.

[0019] As a preferred implementation, the specific steps for performing hierarchical modeling and relationship extraction of business object information to generate a device-process-function relationship network are as follows:

[0020] By reading the fields of equipment number, equipment type, production line, and workstation location of each piece of equipment, the equipment number is used as a unique identifier, and the equipment type, production line, and workstation are used as node attributes to generate a set of equipment-level nodes. By reading the process name, process number, sequence, and key quality control points of each process, the process number is used as a unique identifier, and the process name and sequence information are used as attributes to generate a set of process-level nodes. By reading the function number, function name, system, and call entry point of each functional module, the function number is used as a unique identifier, and the function name and interface or interface address are used as node attributes to generate a set of functional-level nodes.

[0021] As a preferred implementation, the specific steps of performing hierarchical modeling and relationship extraction of business object information to generate a device-process-function relationship network further include:

[0022] By extracting the relationship between the "equipment-process" configuration field in the equipment ledger and the process route, edge relationships are established between equipment in the same process and the corresponding process layer nodes. The workstation number and main purpose information are recorded on the edge, forming a connection from equipment to process. By extracting the fields of "process code", "business scenario" and "interface binding" in the process configuration and function configuration, edge relationships are established between the functions called in a certain process or used for data display and control in that process and the corresponding process layer nodes. The calling direction and calling frequency are recorded on the edge, forming a connection from process to function. The "involved equipment", "involved process" and "involved function" tags in the standardized text set are validated by cross-validation. By storing the equipment layer nodes, process layer nodes, function layer nodes and their edge relationships in the form of graph structure or adjacency list, a cross-domain relationship network of equipment-process-function is generated.

[0023] As a preferred implementation, the specific steps for performing path traversal and weighted summarization on text-related nodes, calculating the impact score of each software function, and generating a score set are as follows:

[0024] The software function impact score set is obtained by locating nodes in a relational network using the "equipment involved," "process involved," and "function involved" tags of each standardized text. The located nodes are designated as the starting nodes of the text, and the text importance is used as the text weight. A finite-depth traversal and path feature calculation method is used to determine the reachable paths from the starting node to each functional layer node. The path weight is obtained by combining the path length, the importance of the nodes traversed, and the reliability of the edge relationships according to a preset formula. The impact score of each functional node is generated by multiplying the path weights with the corresponding text weights, summing them by functional node, and then normalizing the results. This software function impact score set is then provided to subsequent task generation steps. The formula for the propagation path impact score algorithm is as follows:

[0025] ,

[0026] Where IE represents the impact score of the software function node, i is the index of the software function node, T is the current batch of comprehensive data set, and e is any event text in set T. The text weight of event e, Let P be the set of all valid propagation paths from the node aligned to event e to software function i, where p is a specific propagation path from the event-aligned node to the software function node. Let e ​​be the risk coefficient. Let p be the propagation weight. For summation, The dot product symbol;

[0027] The formula for calculating path propagation weight is:

[0028] ,

[0029] in Let be the propagation weight of path p, where p is a specific propagation path from the event alignment node to the software function node, and exp is the exponential function operator. This is the path attenuation coefficient. Let p be the length of path p, and U be the path length. The importance coefficient of node v on the path. Let be the confidence coefficient of the edge from node u to node v in the path, where v represents the target node of the corresponding directed edge, and u represents the starting node of a directed edge in the equipment-process-function relationship network. This is the chain multiplication operator. This is the dot product symbol.

[0030] As a preferred implementation, the specific steps for generating a development task set by performing task template filling and historical case retrieval on the scoring set are as follows:

[0031] By threshold-screening and sorting the impact scores of each software function, functions with scores not lower than a preset threshold are identified as the target function set, and the suggested priorities are sorted from high to low scores. By aggregating information from the standardized text records and equipment-process-function relationship network associated with each target function, the problem phenomenon, triggering scenario, involved equipment, involved process, and scope of impact elements in the functional relationship network are extracted, and preset development task template fields are filled into the extracted elements to generate initial development task entries containing task title, problem description, scenario description, associated objects, and priority.

[0032] As a preferred embodiment, the specific steps of generating a development task set by performing task template filling and historical case retrieval on the scoring set further include:

[0033] Similar event retrieval is performed on historical development tasks and processing records. Using an AI model, records similar to the current task in terms of equipment, process, function, and problem type are selected from the historical database. Processing measures, precautions, and verification points are extracted and added to the corresponding task entry's processing suggestion field. By structuring and encoding the generated development task entries, the tasks are synchronized to the requirements and defect management system, enabling the R&D team to directly receive and track them. Based on a scoring set, a development task generation algorithm is constructed to solve the problems of existing technologies requiring R&D personnel to manually read large amounts of text, rely on experience to judge the scope of impact, and manually write task descriptions. This improves task generation efficiency and the objectivity of decision-making. The formula for the development task generation algorithm is:

[0034] ,

[0035] in Let F be the priority score for the j-th development task, which serves as the direct basis for task sorting and resource allocation, and j be the sequence number of the development task. The weighting coefficients for the normalization results of the functional impact score. This is the normalized result of the software function impact score corresponding to the j-th development task. These are the weighting coefficients of the coverage factor. Let the coverage factor be the j-th development task. These are the weighting coefficients for the urgency factor. Let j be the urgency factor of the j-th development task. To implement the weighting coefficients of cost factors, Let i be the implementation cost factor for the j-th development task, and let i be the index of the software function node.

[0036] As a preferred implementation, the specific steps for forming a self-learning closed-loop model by evaluating the results of the development task after it goes live and updating the parameters are as follows:

[0037] By tracking each generated and issued development task and connecting with the requirements and defect management system, we collect operational result data such as defect reproduction, number of newly added related text events, changes in alarm frequency, and changes in downtime after the corresponding function goes live. We establish a one-to-one correspondence with the software function impact score when the task is generated, and then use rules to judge the above operational results and label the judgment results.

[0038] As a preferred implementation, the specific steps for forming a self-learning closed-loop model by evaluating the results of the development task after it goes live and updating the parameters further include:

[0039] By performing statistical analysis and parameter optimization on the accumulated "resolved", "partially resolved", and "migration risk" labeled samples, the semantic extraction confidence threshold, path decay coefficient, and weights of various elements of the AI ​​big model are adjusted, and the optimization results are written into the model configuration to generate an updated set of data processing model parameters. The updated AI model parameters are then loaded into the subsequent processing flow of the comprehensive data set.

[0040] Beneficial effects

[0041] 1. By uniformly collecting and standardizing multi-source text such as chat logs, work orders, and maintenance logs, and utilizing AI large-scale models for semantic parsing and element extraction, key information such as equipment, processes, functions, and problem types is automatically converted into structured tags, resulting in a standardized text set with semantic tags. This process solves the problems of scattered front-line feedback, inconsistent expression, and difficulty in forming a unified data view caused by relying on keywords and manual processing in existing technologies, providing a stable data foundation for subsequent relationship modeling and quantitative analysis.

[0042] 2. By constructing a cross-domain relationship network consisting of the equipment layer, process layer, and functional layer, and employing an impact scoring algorithm based on propagation path traversal and weighted aggregation on this network, the risk of text events, path length, node importance, and relationship credibility are uniformly quantified into a software function impact score. This technical solution solves the problem in existing technologies that can only perform statistics within a single system and cannot quantify the scope and degree of problem propagation along the "equipment → process → software function" link. This enables software functions to have clear and calculable technical indicators for their impact on the production scenario.

[0043] 3. After obtaining the software function impact score, the priority scoring algorithm comprehensively considers the degree of function impact, cross-factory coverage, risk urgency, and implementation cost to automatically generate a set of structured development tasks with priorities and handling suggestions. The defect reproduction status after software launch and the score changes are used to update the AI ​​model parameters. This mechanism solves the problems in existing technologies where development tasks rely on human experience judgment, priority standards are not uniform, and it is difficult to continuously optimize based on actual results. Thus, a closed loop is achieved from on-site events to task sorting and then to model updates, improving the accuracy and maintainability of the system in long-term operation. Attached Figure Description

[0044] Figure 1 This is a flowchart of the present invention.

[0045] Figure 2 This is a comparison chart of the technical effects of the present invention, in which the black bars represent the present invention and the gray bars represent the prior art. Detailed Implementation

[0046] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.

[0047] Example 1 combined Figure 1 The flowchart shown below illustrates a data processing method for software development applications based on a large AI model. The specific implementation steps are as follows:

[0048] Step S1: Collect frontline feedback and work order data in a unified manner and perform text standardization processing to generate a comprehensive data set;

[0049] Step S1 is used to collect, clean and standardize the format of first-hand feedback and work order texts scattered in various channels such as chat tools, email, and work order system, and form a comprehensive data set that can be processed uniformly.

[0050] Specifically, this is achieved by adopting a unified approach to collecting and standardizing the text of frontline feedback and work order data. Specifically, this involves configuring data collection interfaces between chat tools, email systems, work order systems, and the data platform, and retrieving data containing fields such as problem description, equipment number, process steps, occurrence time, and responsible person at preset time intervals. The collected text is then encoded in a unified format, removing emoticons and meaningless whitespace. Long sentences are broken down and labeled with data source and timestamps line by line. Following a unified data model, the text is mapped to standard fields such as problem description, scenario information, object identifier, and attachment references. Missing fields are filled with placeholders, and the original data primary key is recorded. The processed multi-source data is then written into the same dataset, forming a comprehensive data set that can be used for subsequent semantic analysis.

[0051] Step S2: Based on the comprehensive data set, semantic parsing is performed using an AI large model to extract elements related to business objects and problem types, and generate a standardized text set with labels.

[0052] Step S2 is used to perform semantic analysis on the comprehensive dataset using a large AI model, automatically extracting elements such as equipment, process, function, and problem type from the text to form a standardized text set with semantic labels. The semantic analysis and element extraction algorithm in this step solves the problem in the existing technology that it is difficult to accurately identify equipment, process, function, and problem type information from colloquial and multi-format texts by relying on keywords and fixed rules.

[0053] Specifically, the comprehensive dataset obtained in step S1 is used to perform semantic parsing and element extraction on a pre-trained AI model. This involves constructing a semantic extraction template for industrial manufacturing scenarios, defining the elements to be identified as structured slots such as equipment identifiers, process steps, software functions, problem types, impact scope, and occurrence scenarios. Example expressions and mandatory rules are configured for each slot. Each text in the comprehensive dataset is input into the pre-trained AI model in the form of "contextual description + extraction instructions," enabling the output of each slot's content in a structured format. The output is then parsed to obtain candidate equipment names, process names, function names, and problem type labels. For example, from the text "Welding station A1 frequently alarmed during the night shift yesterday, weld seams..." In the case of "poor welding, suspected unstable parameter distribution", the equipment name is extracted as "Welding Machine A1", the process name is "Welding Process", the function name is "Welding Parameter Distribution Function", and the problem type is "poor welding defect". Based on the pre-maintained equipment ledger, process dictionary and function list, similarity matching and alias mapping are performed on candidate names to uniformly map colloquial abbreviations and spelling errors to standard codes. For missing or conflicting slots, rules are used to complete them according to timestamp, responsible person and system, and confidence level is recorded. By writing the elements extracted and validated by the AI ​​model into the corresponding fields, they are stored together with the original text in a standardized text record to form a standardized text set with semantic tags, which can be called in subsequent relationship network construction and influence scoring steps.

[0054] The semantic extraction template for industrial manufacturing scenarios refers to a set of slot structures and example expressions pre-designed based on the equipment name, process name, software function, and common problem types of industrial manufacturing enterprises. It is used to constrain the AI ​​large model to output structured results according to fields such as equipment identification, process link, software function, problem type, and scope of impact.

[0055] The example expressions and mandatory rules are pre-configured constraint information for each semantic slot in the semantic extraction template. The example expressions are used to list the typical writing and common expressions of equipment identification, process steps, software functions, and problem types in actual text, which are used to guide the AI ​​model to output the corresponding content in the expected format. The mandatory rules are used to indicate whether each slot is a required field and how to handle missing fields. For example, equipment identification is specified as a mandatory field and problem type is a priority field to fill. When the extraction result is missing, the rule is triggered to complete or prompt for supplementation, thereby ensuring the completeness and format consistency of the extraction result.

[0056] The pre-trained AI model refers to a natural language processing model that has been pre-trained on general corpora and industrial scenario corpora and fine-tuned by instructions. It is used to receive input containing on-site text and extraction instructions, and generate structured output or semantic vector representation containing the content of each semantic slot.

[0057] The aforementioned pre-maintained equipment ledger, process dictionary, and function list refer to the equipment master data, process name and code table, and software function list extracted and regularly updated by the enterprise from its existing information system. These are used to provide standard names and unique codes for the large model extraction results, realizing the mapping from natural language expression to standard identifiers.

[0058] The aforementioned similarity matching and alias mapping refers to the method of calculating vector similarity and matching strings between the candidate equipment names, process names, and function names output by the large model and the standard names in the ledger, dictionary, and list, and mapping colloquial abbreviations, abbreviations, or spelling errors to the corresponding standard names and codes according to the preset alias table;

[0059] The rule completion mentioned above refers to a processing method that fills in missing fields or selects and corrects conflict results according to preset rules such as timestamp, responsible person, factory, and default object when some slots are missing or conflicting, and marks the completion results with confidence level.

[0060] Step S3: Based on a standardized text set, a device-process-function relationship network is generated by performing hierarchical modeling and relationship extraction on business object information.

[0061] Step S3 is used to construct the equipment layer, process layer, and function layer and their relationships based on equipment information, process flow and software function configuration, forming a cross-domain relationship network of equipment-process-function.

[0062] Specifically, based on a standardized text set, hierarchical modeling and relationship extraction are performed on equipment, process, and software functional information. Specifically: by reading the equipment number, equipment type, production line, and workstation location of each piece of equipment, the equipment number is used as a unique identifier, and the equipment type, production line, and workstation are used as node attributes to generate an equipment-level node set; by reading the process name, process number, sequence, and key quality control points of each process step, the process number is used as a unique identifier, and the process name and sequence information are used as attributes to generate a process-level node set; by reading the function number, function name, system to which it belongs, and call entry point of each functional module, the function number is used as a unique identifier, and the function... Name, interface, or interface address are used as node attributes to generate a set of functional layer nodes. Based on this, by extracting the relationship between the "equipment-process" configuration field between the equipment ledger and the process route, edge relationships are established between equipment in the same process and the corresponding process layer nodes. Information such as workstation number and main purpose is recorded on the edge, thus forming the association from equipment to process. By extracting fields such as "process code," "business scenario," and "interface binding" from the process configuration and functional configuration using the relationship extraction method, edge relationships are established between the functions called in a certain process or used for data display and control in that process and the corresponding process layer nodes. The calling direction and calling frequency are recorded on the edge, thus forming the association from process to function.

[0063] The aforementioned functional modules are logical units that implement specific business functions in production management or equipment operation and maintenance software. These include modules such as data acquisition, data display, parameter configuration, alarm management, and report generation. They are invoked by users or other systems through the call entry interface to complete the corresponding business processing flow.

[0064] By cross-validating the "Involved Equipment," "Involved Process," and "Involved Function" tags in the standardized text set in step S2, when a text tag is inconsistent with existing relationships or is missing, some edge relationships are added or corrected according to preset rules to complete the implicit associations in actual operation. For example, when the text tag indicates that air compressor A1 experiences pressure fluctuations in the "Compressed Air Supply" process, but the relationship network only has an edge relationship between air compressor A1 and the "Equipment Inspection" process, the system adds an edge between air compressor A1 and the "Compressed Air Supply" process node; when the text tag indicates that "Energy Consumption Early Warning Rule Matching" is not specified, the system adds an edge between air compressor A1 and the "Compressed Air Supply" process node. The "Set" function is used to adjust the alarm threshold of the process. When the call edge between the function and the process is missing in the relationship network, the system adds an edge between the corresponding process layer node and the function layer node and assigns it an initial confidence level to reflect the actual call relationship. By storing the equipment layer node, process layer node, function layer node and their edge relationships in the form of a graph structure or adjacency list, a cross-domain relationship network of equipment-process-function is generated, which provides a unified structured basis for subsequent propagation path traversal and impact score calculation, thereby solving the problem of the lack of a unified relationship model across "equipment-process-software function" in the existing technology.

[0065] Step S4: Based on the functional relationship network, a score set is generated by traversing and weighting the text association nodes and calculating the impact score of each software function.

[0066] Step S4 is used to perform path traversal and weighted calculation by combining text events on the relationship network to obtain the impact score set of each software function and quantify the degree of impact of text events on the function.

[0067] Specifically, by traversing and weighting the propagation path of the cross-domain relationship network, the following steps are taken: The "Involved Equipment," "Involved Process," and "Involved Function" tags of each standardized text obtained in step S2 are used to locate the text within the relationship network using a node positioning method. The located node is taken as the starting node of the text, and the text importance is used as the text weight. A finite-depth traversal and path feature calculation method is used to calculate the reachable paths from the starting node to each functional layer node. The path weight of each path is obtained by combining the path length, the importance of the nodes traversed, and the credibility of the edge relationships according to a preset formula. The impact score of each functional node is generated by multiplying the path weight by the corresponding text weight, accumulating it by functional node, and then normalizing the result. This software function impact score set is then obtained. This software function impact score set is provided to subsequent task generation steps. This propagation path impact scoring algorithm solves the problem in existing technologies that cannot quantify the degree of influence of text events on each software function along the "equipment—process—function" propagation link. The formula for the propagation path impact scoring algorithm is as follows:

[0068] ,

[0069] Where IE is the impact score of the software function node, used to quantify the overall impact of the current set of text events on software function i, serving as the basis for subsequent task generation and priority ranking. i is the index of the software function node, used to uniquely identify a specific software function in the set of function layer nodes, such as the welding parameter distribution interface or the energy consumption early warning rule configuration module. T is the set of texts identified as "events" in the current batch of comprehensive data, representing the overall scope of all text events that need to participate in this impact calculation. e is any event text in set T, used to accumulate the impact of different events on the software function one by one during summation. The text weight of event e is used to reflect the difference in importance between different events. P is the set of all valid propagation paths from the node aligned to event e to software function i, providing the basis for subsequent path-based contribution value calculation. P represents the set of valid propagation paths, where p is a specific propagation path from the event-aligned node to the software function node, used to characterize the specific links in the propagation of event influence within the relational network. Let e ​​be the risk coefficient, used to uniformly represent the risk intensity of the event itself. Let p be the propagation weight, used to represent the effective impact of an event as it propagates through the path to the function. For summation, The dot product symbol;

[0070] The formula for calculating path propagation weight is:

[0071] ,

[0072] in Let be the propagation weight of path p, where p is a specific propagation path from the event alignment node to the software function node, and exp is the exponential function operator. This is the path decay coefficient, a positive constant used to control the rate at which path length diminishes the weight, automatically reducing the contribution of excessively long paths and preventing weakly correlated paths far from the source from having an excessive impact on the score. Let p be the length of the path, and U be the path length used to indicate how many levels the event propagates through. The longer the path, the weaker the direct impact; the shorter the path, the stronger the direct impact. This is the importance coefficient of node v on the path, used to amplify or reduce the impact when passing through critical nodes. For example, the importance coefficient of critical equipment, critical processes, or critical functional nodes is set to a larger value to reflect their amplification effect during propagation. represents the edge reliability coefficient from node u to node v in the path, reflecting the reliability of the relationship between equipment and process, and between process and function. Edges with higher reliability contribute more to the overall impact. v represents the target node of the corresponding directed edge, which can be an equipment node, process node, or function node, used to represent the downstream node on the propagation path. u represents the starting node of a directed edge in the equipment-process-function relationship network, used to represent the upstream node on the propagation path. This is the chain multiplication operator, used to multiply multiple values ​​in a set sequentially. The dot product symbol;

[0073] The propagation path impact scoring algorithm traverses each text event along the reachable path on the equipment-process-function relationship network, and calculates the software function impact score by combining event weight, risk coefficient, path length, node importance and edge credibility. This quantifies "from which equipment, which process, and which function does the problem propagate to, and how great is the impact" into a clear numerical value, solving the problem that existing technologies can only perform simple statistics and cannot quantitatively assess the degree of impact on software functions along the "equipment → process → function" link.

[0074] The above formula unifies and quantifies text risk, path length, node importance, and relationship credibility, generating impact scores for each software function. This effectively solves the problem in existing technologies that cannot accurately compare the degree of impact of software functions on different propagation paths and different factories by relying solely on the frequency of occurrence or a single dimension.

[0075] Step S5: Based on the rating set, generate a development task set by performing task template population and historical case retrieval on the rating set;

[0076] Step S5 is used to automatically generate a set of structured development tasks with priorities and processing suggestions based on software function impact scores and related text, for the R&D team to execute.

[0077] Specifically, this is achieved by generating tasks and outputting optimization suggestions based on the software function impact score set. Specifically, this involves: using a threshold filtering and sorting method for the impact scores of each software function, identifying functions with scores above a preset threshold as the target function set, and prioritizing suggestions according to scores from highest to lowest; aggregating information from standardized text records and equipment-process-function relationship networks associated with each target function, extracting elements such as problem phenomena, triggering scenarios, involved equipment, involved processes, and scope of impact from the functional relationship network, and filling the extracted elements with preset development task template fields to generate initial development task entries containing task titles, problem descriptions, scenario descriptions, associated objects, and priorities; and reviewing historical development task entries. The system retrieves similar events from task assignment and processing records. Using an AI model, it selects several records from the historical database that are similar to the current task in terms of equipment, process, function, and problem type. It then extracts the processing measures, precautions, and verification points from these records and adds them to the corresponding task entry's processing suggestion field. By structurally encoding the generated development task entries, the tasks are synchronized to the requirements and defect management system, enabling the R&D team to directly receive and track them. Based on a scoring set, a development task generation algorithm is constructed to solve the problems of existing technologies requiring R&D personnel to manually read large amounts of text, rely on experience to judge the scope of impact, and manually write task descriptions. This improves task generation efficiency and the objectivity of decision-making. The formula for the development task generation algorithm is as follows:

[0078] ,

[0079] in Let F be the priority score for the j-th development task. The priority score quantifies the urgency and importance of this task within the current batch of tasks, serving as a direct basis for task sorting and resource allocation. j is the sequence number of the development task, used to uniquely identify a specific task within the development task set, such as adjusting welding parameter distribution logic or adding an energy consumption comparison report. These are weighting coefficients for the normalized results of the functional impact score, used to control the proportion of the degree of impact of software functional technology in the development task priority score. This is a normalized result for the software function impact score corresponding to the j-th development task, used to convert the impact scores of different functions into dimensionless values ​​between 0 and 1, facilitating weighted summation with other factors. This is a weighting factor for the coverage factor, used to control the proportion of the factory or production line scope involved in the task in the priority score. Let be the coverage factor for the j-th development task, representing the scope of the factory, production line, or business unit involved in that task. This is the weighting coefficient for the urgency factor, used to control the proportion of safety risk, compliance risk, and production disruption risk levels in the priority score. Let be the urgency factor for the j-th development task. It is obtained by mapping information such as the level of security risks, compliance risks, and production interruption risks involved in the task, and is used to characterize the urgency of processing the task in the time dimension. The weighting coefficients for cost factors are used to control the proportion of cost elements such as development workload, testing workload, and downtime in the priority scoring. Let i be the implementation cost factor for the j-th development task, used to reflect the priority value of "low cost and high return", and i be the index of the software function node.

[0080] By weighting the impact score normalization result, coverage factor, urgency factor and implementation cost factor, a priority score for each development task is obtained. The technical impact, scope, risk urgency and implementation cost are uniformly incorporated into a quantitative indicator, which solves the problem that the priority of development tasks in the existing technology mainly relies on subjective judgment based on human experience and that the standards of different teams are inconsistent.

[0081] The aforementioned requirement and defect management system is an information system used to record, track, and manage software requirement items and defect items. It is used to receive the structured development task set generated by this invention, and to uniformly register and track the requirement number, defect number, status transition, and processing result corresponding to the task, so as to provide task execution process data for subsequent online effect evaluation and model update.

[0082] Step S6: By evaluating the results of the development task after it goes online and updating the parameters, a self-learning closed-loop model is formed.

[0083] Step S6 is used to collect the running effect after the development task is launched, and adjust the model parameters according to the defect reproduction and score changes to form a self-learning closed-loop model, thereby improving the accuracy of subsequent identification and recommendation.

[0084] Specifically, each development task generated and issued in step S5 is tracked and integrated with the requirements and defect management system. Data on defect reproduction, the number of newly added related text events, changes in alarm frequency, and changes in downtime are collected after the corresponding function goes live. A one-to-one correspondence is established between these data and the software function impact score at the time the task was generated. Rule-based judgments are applied to these results, and the judgment results are marked. When no similar or identical problems appear during the observation period and the impact score of related text events decreases significantly, the corresponding task is marked as "resolved." For example, after the development task for "welding parameter issuance function causing weld joint defects" went live, no similar quality complaints appeared within a one-month observation period, and the impact score of text events related to the welding station decreased from 0.85 to 0.10. When problems still occur sporadically but the impact score has decreased, it is marked as "partially resolved." For example, regarding the issue of "low energy consumption warning threshold causing frequent..." After the development task of "false alarms" was launched, the number of energy consumption false alarms decreased significantly but still occurred occasionally, and the impact score of related text events decreased from 0.90 to 0.40. When the false alarms reappeared in other factories or production lines along a similar "equipment-process-function" path, they were marked as "migration risk" and the characteristics of the new propagation path were recorded. Statistical analysis and parameter optimization were performed on the accumulated "resolved", "partially resolved", and "migration risk" labeled samples. By adjusting the semantic extraction confidence threshold in step S2, the path decay coefficient in step S4, and the weights of various elements in the AI ​​large model, and writing the optimization results into the model configuration, an updated set of data processing model parameters was generated. By loading the updated AI model parameters into the subsequent processing flow of the comprehensive data set, a self-learning closed loop based on the launch effect was achieved, solving the problem that the existing technology model relies on static rules for a long time and cannot automatically improve the problem recognition rate and task recommendation accuracy by combining processing effects.

[0085] Combination Figure 2 The image shows a comparison of the technical effects of a data processing method for software development applications based on AI large-scale models. The black bars represent the technical effects of the present invention, while the gray bars represent the technical effects of existing technologies. Figure 2 It can be seen that the technical effect of the present invention is superior to that of the prior art.

[0086] Example 2, based on Example 1, is a data processing method for software development applications using a large AI model. The specific solution is as follows:

[0087] In this embodiment, the company has deployed an energy management system, a MES system, and an operation and maintenance system in factories A, B, and C, involving energy-consuming equipment such as air compressors, refrigeration units, and heat treatment furnaces. Frontline energy management personnel report issues such as "abnormal peak electricity consumption during a certain period," "suspected compressed air leakage," and "low efficiency of the chilled water system" through team WeChat groups and the energy consumption work order system. The energy management system and equipment monitoring system generate text data including alarm descriptions, abnormal power curve descriptions, and energy consumption report notes.

[0088] Step S1: By adopting a unified collection and text standardization method for energy consumption work orders, team chat records, energy consumption report notes and operation and maintenance records of each factory, the descriptive information from different factories and different systems is cleaned, time-aligned and labeled with factory identifiers to generate a cross-factory comprehensive data set.

[0089] Step S2: Based on the comprehensive data set, semantic parsing and element extraction are performed using an AI big data model to extract equipment names (e.g., air compressor A1, refrigeration unit B2), process links (e.g., compressed air supply, cooling water circulation), software functions (e.g., real-time energy consumption monitoring interface, energy consumption early warning rule configuration module), problem types (e.g., peak limit exceeded, efficiency decline, suspected leakage), and scope of impact (number of factories and production lines involved) from the text, generating a standardized text set with semantic tags;

[0090] Step S3: Based on a standardized text set, a hierarchical modeling and relationship extraction method is adopted for the equipment ledgers, energy process flow and energy management software function configuration tables of each factory. Energy consumption equipment nodes of each factory are established at the equipment layer, nodes such as compressed air supply, cooling water circulation and heat treatment process are established at the process layer, and software function nodes such as energy consumption monitoring, load forecasting, energy consumption report generation and energy consumption early warning rules are established at the function layer. A cross-domain relationship network of equipment-process-function is constructed according to the equipment affiliation, process usage relationship and function call relationship.

[0091] Step S4: When a cross-factory energy consumption anomaly occurs during a certain period, the standardized text records related to the event are located by node positioning, and the propagation path from the event-related node to each energy consumption-related function node is traversed and weighted and summarized on the cross-domain relationship network according to the software function impact scoring algorithm described in Example 1. The impact score of each energy consumption function is calculated to form an energy consumption function impact score set, which is used to identify the criticality of functions such as "energy consumption early warning rule configuration module" and "energy consumption monitoring interface" in this anomaly.

[0092] Step S5: Based on the energy consumption function impact score set, by performing task template filling and historical similar event retrieval on function nodes with high impact scores and large coverage factors, a structured development task set for energy consumption scenarios is generated, such as "adjusting the compressed air early warning threshold of a factory", "adding a cross-factory energy consumption comparison view", and "supplementing the peak cause field in the energy consumption report". Then, according to the development task priority scoring algorithm described in Example 1, priority scores are calculated for each task to obtain an energy consumption optimization task list with quantified priority and processing suggestions.

[0093] Step S6: After the corresponding software version is launched, by monitoring the number of energy consumption anomalies, the proportion of peak exceedances, and the changes in the impact scores of related text events in each factory, the energy consumption optimization tasks generated in this embodiment are labeled with tags such as "resolved," "partially resolved," or "migration risk." Based on these label samples, the path decay coefficient and node weight in the function impact scoring algorithm, as well as the weight coefficient in the development task priority scoring algorithm, are adjusted to make the subsequent identification and task ranking for energy consumption anomaly scenarios more consistent with actual results.

[0094] This embodiment demonstrates that the AI-based large model-based software development application data processing method proposed in this invention is not only applicable to the single-factory welding quality problem scenario in Embodiment 1, but also to the multi-factory energy consumption anomaly and compliance management scenario in this embodiment, enabling software function impact assessment and development task priority ranking under cross-factory and consistent standards.

[0095] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A data processing method for software development applications based on AI large-scale models, characterized in that: Step S1: Collect frontline feedback and work order data in a unified manner and perform text standardization processing to generate a comprehensive data set; Step S2: Based on the comprehensive data set, semantic parsing is performed using an AI large model to extract elements related to business objects and problem types, and generate a standardized text set with labels. Step S3: Based on a standardized text set, a device-process-function relationship network is generated by performing hierarchical modeling and relationship extraction on business object information. Step S4: Based on the functional relationship network, a score set is generated by traversing and weighting the text association nodes and calculating the impact score of each software function. Step S5: Based on the rating set, generate a development task set by performing task template population and historical case retrieval on the rating set; Step S6 involves evaluating the performance and updating the parameters of the development task after it goes live, thus forming a self-learning closed-loop model.

2. The software development application data processing method based on a large AI model according to claim 1, characterized in that: The specific steps for uniformly collecting frontline feedback and work order data, performing text standardization processing, and generating a comprehensive dataset are as follows: By configuring data collection interfaces between chat tools, email systems, work order systems, and data platforms, and retrieving field data containing problem descriptions, equipment numbers, process steps, occurrence times, and responsible persons at preset time intervals, the collected text data is encoded in a unified format, removing emojis and meaningless whitespace. Long sentences are broken down and labeled with data sources and timestamps line by line. Following a unified data model, the text is mapped to standard fields for problem descriptions, scenario information, object identifiers, and attachment references. Missing fields are filled with placeholders, and the original data primary keys are recorded. The processed multi-source data is written into the same dataset, forming a comprehensive data set that can be used for subsequent semantic analysis.

3. The software development application data processing method based on a large AI model according to claim 1, characterized in that: The specific steps for using a large AI model to perform semantic parsing, extract elements related to business objects and problem types, and generate a standardized text set with labels are as follows: By constructing a semantic extraction template for industrial manufacturing scenarios, the elements to be identified are defined as structured slots such as equipment identifier, process link, software function, problem type, scope of impact, and occurrence scenario. Example expressions and mandatory rules are configured for each slot. Each text in the comprehensive dataset is input into a pre-trained AI model in the form of "context description + extraction instructions". The output results are parsed to obtain candidate equipment names, process names, function names, and problem type labels. Based on pre-maintained equipment ledgers, process dictionaries, and function lists, similarity matching and alias mapping are performed on candidate names to uniformly map colloquial abbreviations and spelling errors to standard codes. For missing or conflicting slots, rules are used to complete them according to timestamps, responsible persons, and systems, and confidence level markers are recorded. By writing the elements extracted and validated by the AI ​​model into the corresponding fields, they are stored together with the original text in standardized text records to form a standardized text set with semantic tags.

4. The software development application data processing method based on a large AI model according to claim 1, characterized in that: The specific steps for performing hierarchical modeling and relationship extraction of business object information to generate a device-process-function relationship network are as follows: By reading the fields of equipment number, equipment type, production line, and workstation location of each piece of equipment, the equipment number is used as a unique identifier, and the equipment type, production line, and workstation are used as node attributes to generate a set of equipment-level nodes. By reading the process name, process number, sequence, and key quality control points of each process, the process number is used as a unique identifier, and the process name and sequence information are used as attributes to generate a set of process-level nodes. By reading the function number, function name, system, and call entry point of each functional module, the function number is used as a unique identifier, and the function name and interface or interface address are used as node attributes to generate a set of functional-level nodes.

5. The software development application data processing method based on a large AI model according to claim 4, characterized in that: The specific steps for performing hierarchical modeling and relationship extraction of business object information to generate a device-process-function relationship network also include: By extracting the relationship between the "equipment-process" configuration field in the equipment ledger and process route, edge relationships are established between equipment in the same process and corresponding process layer nodes, and the workstation number and main purpose information are recorded on the edge to form the association from equipment to process. By extracting the fields of "process code", "business scenario" and "interface binding" in process configuration and function configuration, edge relationships are established between the functions called in a certain process or used for data display and control in that process and corresponding process layer nodes, and the calling direction and calling frequency are recorded on the edge to form the association from process to function. The "involved equipment", "involved process" and "involved function" tags in the standardized text set are validated by cross-validation. By storing the equipment layer nodes, process layer nodes, function layer nodes and their edge relationships in the form of graph structure or adjacency list, a cross-domain relationship network of equipment-process-function is generated.

6. The software development application data processing method based on a large AI model according to claim 1, characterized in that: The specific steps for performing path traversal and weighted summarization of text-related nodes, calculating the impact score of each software function, and generating a score set are as follows: The software function impact score set is obtained by locating nodes in a relational network using the "Involved Equipment," "Involved Process," and "Involved Function" tags of each standardized text. The located nodes are designated as the starting nodes of the text, and the text importance is used as the text weight. A finite-depth traversal and path feature calculation method is employed to determine the reachable paths from the starting node to each functional node. The path weight is calculated by combining path length, node importance, and edge relationship reliability according to a preset formula. The impact score of each functional node is generated by multiplying the path weights with their corresponding text weights, summing them by functional node, and then normalizing the results. This software function impact score set is then provided to subsequent task generation steps. The formula for the propagation path impact score algorithm is as follows: ， Where IE represents the impact score of the software function node, i is the index of the software function node, T is the current batch of comprehensive data set, and e is any event text in set T. The text weight of event e, Let P be the set of all valid propagation paths from the node aligned to event e to software function i, where p is a specific propagation path from the event-aligned node to the software function node. Let e ​​be the risk coefficient. Let p be the propagation weight. For summation, The dot product symbol; The formula for calculating path propagation weight is: ， in Let be the propagation weight of path p, where p is a specific propagation path from the event alignment node to the software function node, and exp is the exponential function operator. This is the path attenuation coefficient. Let p be the length of path p, and U be the path length. The importance coefficient of node v on the path. Let be the confidence coefficient of the edge from node u to node v in the path, where v represents the target node of the corresponding directed edge, and u represents the starting node of a directed edge in the equipment-process-function relationship network. This is the chain multiplication operator. This is the dot product symbol.

7. The software development application data processing method based on a large AI model according to claim 1, characterized in that: The specific steps for generating a development task set by performing task template population and historical case retrieval on the rating set are as follows: By threshold-screening and sorting the impact scores of each software function, functions with scores not lower than a preset threshold are identified as the target function set, and the suggested priorities are sorted from high to low scores. By aggregating information from the standardized text records and equipment-process-function relationship network associated with each target function, the problem phenomenon, triggering scenario, involved equipment, involved process, and scope of impact elements in the functional relationship network are extracted, and preset development task template fields are filled into the extracted elements to generate initial development task entries containing task title, problem description, scenario description, associated objects, and priority.

8. The software development application data processing method based on a large AI model according to claim 7, characterized in that: The specific steps for generating a development task set by performing task template population and historical case retrieval on the rating set also include: Similar event retrieval is performed on historical development tasks and processing records. Using an AI model, records similar to the current task in terms of equipment, process, function, and problem type are selected from the historical database. Processing measures, precautions, and verification points are extracted and added to the corresponding task entry's processing suggestion field. By structuring and encoding the generated development task entries, the tasks are synchronized to the requirements and defect management system, enabling the R&D team to directly receive and track them. Based on a scoring set, a development task generation algorithm is constructed to solve the problems of existing technologies requiring R&D personnel to manually read large amounts of text, rely on experience to judge the scope of impact, and manually write task descriptions. This improves task generation efficiency and the objectivity of decision-making. The formula for the development task generation algorithm is: ， in Let F be the priority score for the j-th development task, which serves as the direct basis for task sorting and resource allocation, and j be the sequence number of the development task. The weighting coefficients for the normalization results of the functional impact score. This is the normalized result of the software function impact score corresponding to the j-th development task. These are the weighting coefficients of the coverage factor. Let the coverage factor be the j-th development task. These are the weighting coefficients for the urgency factor. Let j be the urgency factor of the j-th development task. To implement the weighting coefficients of cost factors, Let i be the implementation cost factor for the j-th development task, and let i be the index of the software function node.

9. The software development application data processing method based on a large AI model according to claim 1, characterized in that: The specific steps for forming a self-learning closed-loop model by evaluating the results of the development task after it goes live and updating the parameters are as follows: By tracking each generated and issued development task and connecting with the requirements and defect management system, we collect operational result data such as defect reproduction, number of newly added related text events, changes in alarm frequency, and changes in downtime after the corresponding function goes live. We establish a one-to-one correspondence with the software function impact score when the task is generated, and then use rules to judge the above operational results and label the judgment results.

10. The software development application data processing method based on a large AI model according to claim 9, characterized in that: The specific steps for forming a self-learning closed-loop model by evaluating the results of the development task after it goes online and updating the parameters also include: By performing statistical analysis and parameter optimization on the accumulated "resolved", "partially resolved", and "migration risk" labeled samples, the semantic extraction confidence threshold, path decay coefficient, and weights of various elements of the AI ​​large model are adjusted, and the optimization results are written into the model configuration to generate an updated set of data processing model parameters. The updated AI model parameters are then loaded into the subsequent processing flow of the comprehensive data set.

Images