Process parameter optimization diagnosis method fusing large language model and structured model

Abstract

The invention discloses a process parameter optimization diagnosis method fusing a large language model and a structured model, and aims to improve defect identification precision and diagnosis interpretation capability in a welding process. The method comprises four steps of process parameter acquisition and preprocessing, welding defect identification model construction, structured output alignment and language model reasoning, and man-machine interaction and feedback display. Firstly, various technological parameters in the welding process are collected, and standardization and sample balance treatment are carried out; secondly, a structured diagnosis model is constructed, and classified recognition of welding defects such as air holes and tungsten inclusion is achieved; converting structured output into natural language prompts, inputting the natural language prompts into the fine-tuned large language model, and performing defect cause interpretation and process optimization suggestion generation; and finally, displaying a diagnosis result and visual feedback to a user through a natural language interaction interface. The method has high accuracy and interpretability, and is suitable for intelligent manufacturing and industrial welding quality monitoring scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent manufacturing and industrial artificial intelligence, in particular to a process parameter optimization diagnosis method combining a large language model and a structured model, which is suitable for multi-parameter driven welding defect identification and man-machine interactive intelligent diagnosis scenarios. BACKGROUND

[0002] Welding, as a key connection process, plays an important role in high-end manufacturing industries such as aerospace, automobile manufacturing, and shipbuilding. Traditional welding quality detection methods mainly rely on manual experience, rule systems, or image recognition algorithms, which have limited classification accuracy, weak defect interpretation ability, and lack of interactive feedback, making it difficult to meet the industrial application requirements under complex working conditions and high quality standards. In recent years, structured data-driven welding defect classification models have made some progress, such as XGBoost and deep neural networks showing good performance in multi-variable parameter modeling. At the same time, large language models have shown significant advantages in natural language understanding and task reasoning. However, existing methods generally have problems such as separation of structured models and language models, incomplete reasoning links, poor parameter interpretability, and insufficient interaction capabilities, which restrict the further development of welding intelligent diagnosis systems. SUMMARY

[0003] To solve the above problems, the present application proposes a process parameter optimization diagnosis method combining a large language model and a structured model, which is used to solve the deficiencies in defect identification accuracy, reasoning ability, semantic understanding, and interaction ability in existing welding quality detection systems. Through a prompt alignment mechanism and a function call framework, deep collaboration between structured data and language models is achieved, supporting multi-task closed-loop reasoning such as defect identification, cause analysis, and parameter optimization suggestions. This method has good interpretability, scalability, and human-computer interaction ability while ensuring high diagnostic accuracy, and is suitable for various industrial welding quality evaluation and intelligent decision-making scenarios.

[0004] The technical solution adopted by the present application to achieve the above-mentioned purposes is as follows:

[0005] The process parameter optimization diagnosis method combining a large language model and a structured model includes the following steps:

[0006] A process data acquisition and preprocessing module acquires and processes process parameter data during welding;

[0007] A welding defect identification model construction module uses a welding defect identification model to identify and predict welding defects;

[0008] A structured output alignment and language model diagnosis module aligns the structured prediction results to natural language prompts and performs diagnosis reasoning through a language model;

[0009] The interactive interface and intelligent feedback display module displays the diagnosis result and collects user feedback information.

[0010] The process data acquisition and preprocessing module performs the following steps:

[0011] The parameter acquisition submodule acquires six-dimensional parameters in the welding process;

[0012] The data cleaning submodule performs outlier rejection and missing value filling on the six-dimensional parameters;

[0013] The feature normalization submodule performs standardization or normalization processing on the cleaned data;

[0014] The class balancing submodule performs sample enhancement on the normalized data;

[0015] The data segmentation and caching submodule divides the enhanced data into a training set, a validation set, and a test set, and performs caching management.

[0016] The welding defect recognition model construction module performs the following steps:

[0017] The feature interaction construction submodule models the coupling relationship between multiple process parameters to obtain interactive features of the process parameters;

[0018] The XGBoost model submodule obtains high-order nonlinear interaction relationships between the interactive features of the process parameters;

[0019] The TabNet model submodule performs dynamic feature selection on the interactive features of the process parameters through an attention mechanism;

[0020] The Stacking fusion submodule fuses the output results of the XGBoost model and the TabNet model based on a meta-learner to obtain a defect prediction label;

[0021] The model evaluation and selection submodule verifies the welding defect recognition model using the validation set and determines the optimal model configuration based on the verification result.

[0022] The structured output alignment and language model diagnosis module performs the following steps:

[0023] The output structure analysis submodule extracts the prediction label of the welding defect recognition model;

[0024] The prompt template construction submodule converts the structured prediction label into a natural language prompt;

[0025] The domain language model inference submodule performs defect reason analysis, parameter suggestion generation, and task chain invocation based on the Qwen2.5-1.5B model;

[0026] The diagnostic suggestion generation submodule outputs process optimization suggestions and coping strategies in natural language format.

[0027] The structured output alignment and language model diagnosis module further performs the following steps:

[0028] The function scheduling submodule triggers the welding defect recognition model, database or external module call according to the large language model judgment or user instruction.

[0029] The field language model inference submodule is a large language model that is pre-trained on the basis of Qwen2.5-1.5B and fine-tuned on welding field corpus.

[0030] The interactive interface and intelligent feedback display module performs the following steps:

[0031] The user input analysis submodule analyzes the query, instruction and dialogue content proposed by the user;

[0032] The multi-round interaction tracking submodule records and processes multi-round interaction information in continuous context;

[0033] The visualization display submodule displays the diagnostic results to the user in the form of charts, text or mixed layout;

[0034] The user feedback collection submodule collects user correction opinions and labeled feedback information;

[0035] The permission control and deployment interface submodule manages user access permissions and supports local / remote deployment docking of the system.

[0036] The process parameter optimization diagnosis system integrating the large language model and the structured model comprises:

[0037] The process data acquisition and preprocessing module is used for acquiring and processing process parameter data in the welding process;

[0038] The welding defect recognition model construction module is used for identifying and predicting welding defects using a welding defect recognition model;

[0039] The structured output alignment and language model diagnosis module is used for aligning the structured prediction results to natural language prompts and performing diagnosis and reasoning through a language model;

[0040] The interactive interface and intelligent feedback display module is used for displaying the diagnostic results and collecting user feedback information.

[0041] The present application has the following advantages and benefits:

[0042] 1. Improve the accuracy of welding defect recognition: the application constructs a hybrid structured model WeldStructNet that combines XGBoost and TabNet, which effectively improves the defect classification performance under multiple variable process parameters by combining the structure modeling capabilities and feature selection mechanisms of the two, and adapts to complex welding conditions.

[0043] 2. Enhance the semantic reasoning ability of the diagnosis system: by introducing a large language model fine-tuned in the welding field, the application supports natural language understanding and task reasoning for structured output, can explain the causes of defects and generate parameter optimization suggestions, and improves the explainability and knowledge expression ability of the system.

[0044] 3. Implement multi-round task interaction and function calling mechanism: the application supports users to propose multi-round natural language instructions through prompt alignment and function calling mechanism, realizes the closed-loop interaction process of model calling, result feedback and task scheduling, and has good human-computer interface friendliness.

[0045] 4. Support cross-model collaboration and scalable deployment: structured models and large language models run collaboratively, and the system structure has modular design, which is convenient for embedded deployment in industrial field, and can also be extended to integrate image, sound and other multi-modal data sources, to build a more comprehensive intelligent welding diagnosis system. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The system architecture diagram of the process parameter optimization diagnosis method of the application combining large language model and structured model;

[0047] Figure 2 The processing flow chart of the structured defect recognition model in the application;

[0048] Figure 3 The interaction logic diagram of the application in which the large language model calls the structured model and generates diagnosis suggestions. DETAILED DESCRIPTION

[0049] The application will be further described in detail below in combination with the drawings and examples.

[0050] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the application belongs. The terms used in the specification of the application are only for the purpose of describing specific embodiments and are not intended to limit the application.

[0051] A process parameter optimization diagnosis method combining a large language model and a structured model, comprising the following steps:

[0052] Six key process parameters in the welding process, including current, voltage, humidity, gas flow, workpiece temperature and ambient temperature, are acquired, and data cleaning, normalization and sample balancing processing are performed to generate training samples available for structured models;

[0053] A welding structured diagnosis model WeldStructNet is constructed by fusing XGBoost and TabNet, the nonlinear interaction relationship and feature attention weight are modeled respectively, and the meta-learner is used to fuse the outputs of multiple models to output the welding defect prediction results and their confidence;

[0054] The output of the structured model (defect label, weight explanation, etc.) is converted into a prompt, which is injected into the fine-tuned large language model to complete the natural language explanation of the welding defect causes, parameter optimization suggestions and external tool calls;

[0055] Users initiate diagnosis requests through a natural language interface, and the system returns defect analysis, adjustment suggestions and auxiliary knowledge reasoning results, supports multi-round interaction, visual text and image output, and user feedback collection for closed-loop optimization and continuous learning of the model.

[0056] As shown in Figure 1 A process parameter optimization diagnosis method combining large language models and structured models mainly includes the following four modules: process data acquisition and preprocessing module, welding defect recognition model construction module, structured output alignment and language model diagnosis module, and human-computer interaction and feedback display module.

[0057] Process data acquisition and preprocessing module: This module is used to collect multi-source process parameter data in the welding process, including six key features such as current, voltage, humidity, gas flow, workpiece temperature and ambient temperature. The collected data is first cleaned by the data cleaning module to remove outliers and fill in missing information, and then normalized algorithm (such as Z-score standardization or MinMax scaling) is used for feature scaling to improve the stability and generalization ability of model training. In addition, considering the problem of unbalanced number of defect categories in industrial scenarios, the SMOTE oversampling algorithm is introduced to balance the training data set, enhancing the learning ability of small sample defect categories.

[0058] The process data acquisition and preprocessing module includes the following sub-modules:

[0059] Parameter acquisition sub-module, six-dimensional parameters such as welding current, voltage, gas flow, humidity, workpiece temperature and ambient temperature are collected;

[0060] Data cleaning sub-module, outlier removal and missing value filling are performed;

[0061] Feature normalization sub-module, standardization or normalization processing is performed;

[0062] Class balance sub-module, using SMOTE and other methods for sample enhancement;

[0063] Data segmentation and caching sub-module, divide data into training, validation, and test sets, and perform cache management.

[0064] Weld defect recognition model construction module: as shown in Figure 2 The module constructs a hybrid structured classification model WeldStructNet that combines XGBoost and TabNet, respectively mining the nonlinear high-order interaction between variables and dynamic feature attention ability. The XGBoost part is based on decision tree ensemble algorithm, modeling the combined effect and boundary division of process parameters; the TabNet part dynamically selects features through sparse attention mechanism, improving the model interpretability and robustness. Finally, the output results of the two sub-models are fused through the meta-learner (Meta Learner), realizing high-precision and multi-classification of welding defects.

[0065] The weld defect recognition model construction module includes the following sub-modules:

[0066] XGBoost model sub-module, modeling high-order nonlinear interaction of process parameters;

[0067] TabNet model sub-module, dynamic feature selection through attention mechanism;

[0068] Feature interaction construction sub-module, combined modeling of the coupling relationship between multiple parameters;

[0069] Stacking fusion sub-module, using ensemble learning strategy to fuse multiple model outputs;

[0070] Model evaluation and selection sub-module, based on the validation results to determine the optimal model configuration.

[0071] The Stacking fusion sub-module is based on the meta-learner to perform secondary training and learning on the output results of the XGBoost and TabNet models, to improve the classification accuracy and model generalization ability of defect recognition.

[0072] Structured output alignment and language model diagnosis module: as shown in Figure 3To achieve seamless connection between structured prediction results and natural language reasoning, the module parses the output results of WeldStructNet (defect labels, confidence, feature importance, etc.) into standardized text descriptions, constructs prompt templates, and injects them as input into the fine-tuned large language model. The large language model has the ability of welding semantic understanding and task execution, and can complete defect cause explanation, parameter optimization suggestion generation, and task tool chain invocation functions in combination with user queries. At the same time, the module supports function-level call scheduling mechanism, and the language model can actively trigger the structure model, knowledge base or other external modules according to the task intent, to realize dynamic reasoning and task linkage.

[0073] The structured output alignment and language model diagnosis module includes the following submodules:

[0074] The output structure parsing submodule extracts the predicted labels and feature importance of the structured model;

[0075] The prompt template construction submodule converts structured information into natural language prompts;

[0076] The domain language model reasoning submodule performs defect cause analysis, parameter suggestion generation, and task chain invocation based on the Qwen2.5-1.5B model;

[0077] The function scheduling submodule triggers the structure model, database, or external module call according to the model judgment or user instruction;

[0078] The diagnosis suggestion generation submodule outputs process optimization suggestions and coping strategies in natural language format.

[0079] The language model is a large language model based on Qwen2.5-1.5B, which is further pre-trained and fine-tuned on welding domain corpus, and has the ability of multi-round task understanding and language interaction execution.

[0080] Human-computer interaction and feedback display module: users can ask the system for queries through natural language interface, such as "why are there air holes" and "how to optimize the current voltage setting". The system generates multi-dimensional diagnostic feedback based on the reasoning results of the structure model and the language model after analyzing the user's intent. This module provides multi-round dialogue mechanism to support context retention, allowing users to continuously ask for reasons or suggestions. Feedback information is displayed in a mixed layout of text and images, including defect type, formation mechanism, parameter adjustment suggestion, etc. The system also supports user feedback on the diagnostic results for subsequent model optimization and continuous learning.

[0081] The interactive interface and intelligent feedback display module includes the following submodules:

[0082] The user input analysis submodule analyzes the user's queries, instructions, and conversation content;

[0083] Multi-round interaction tracking sub-module, records and processes multi-round interaction information in continuous context;

[0084] Visual display sub-module, displays the diagnosis results to the user in the form of charts, words or mixed layout;

[0085] User feedback collection sub-module, collects user's correction opinions and labeled feedback information;

[0086] Permission control and deployment interface sub-module, manages user access permissions and supports local / remote deployment docking of the system.

[0087] Through the above-mentioned modules work together, the present application realizes the closed-loop intelligent diagnosis process from parameter perception, defect identification to language reasoning, suggestion generation, has high accuracy, high interpretability and good interactive experience, and is suitable for various industrial welding process quality monitoring and intelligent auxiliary decision-making tasks.

[0088] The above is the preferred embodiment of the present application, those skilled in the art can make various modifications and improvements to the present application without departing from the spirit and principles of the present application, and these modifications and improvements should also be considered as the protection scope of the present application.

Claims

1. A process parameter optimization and diagnosis method integrating a large language model and a structured model, characterized in that: The following steps are involved: The process data acquisition and preprocessing module acquires and processes the process parameter data during the welding process; The welding defect recognition model building module uses the welding defect recognition model to identify and predict welding defects; The structured output alignment and language model diagnosis module aligns structured prediction results to natural language prompts and performs diagnostic reasoning through language models; The interactive interface and intelligent feedback display module displays the diagnosis results and collects user feedback information.

2. The process parameter optimization diagnosis method integrating a large language model and a structured model according to claim 1 is characterized in that: The process data acquisition and preprocessing module performs the following steps: The parameter acquisition submodule collects six-dimensional parameters during the welding process; The data cleaning submodule removes outliers and fills missing values ​​for the six-dimensional parameters; The feature normalization submodule standardizes or normalizes the cleaned data; The category balancing submodule performs sample enhancement on the normalized data; The data segmentation and caching submodule divides the enhanced data into training set, validation set and test set, and performs cache management.

3. The process parameter optimization diagnosis method integrating a large language model and a structured model according to claim 1 is characterized in that: The welding defect recognition model building module performs the following steps: The feature interaction construction submodule performs combined modeling on the coupling relationship between multiple process parameters to obtain the interaction characteristics of process parameters; The XGBoost model submodule obtains high-order nonlinear interaction relationships between process parameter interaction features; The TabNet model submodule performs dynamic feature selection on the process parameter interaction features through the attention mechanism; The Stacking fusion submodule fuses the output results of the XGBoost model and the TabNet model based on the meta-learner to obtain the defect prediction label; The model evaluation and selection submodule uses the validation set to validate the welding defect recognition model and determines the optimal model configuration based on the validation results.

4. The process parameter optimization diagnosis method integrating a large language model and a structured model according to claim 1 is characterized in that: The structured output alignment and language model diagnosis module performs the following steps: The output structure parsing submodule extracts the predicted labels of the welding defect recognition model; The prompt template construction submodule converts structured prediction labels into natural language prompts; The domain language model reasoning submodule performs defect cause analysis, parameter suggestion generation, and task chain invocation based on the Qwen2.5-1.5B model; The diagnostic suggestion generation submodule outputs process optimization suggestions and response strategies in natural language format.

5. The process parameter optimization diagnosis method integrating a large language model and a structured model according to claim 4 is characterized in that: The structured output alignment and language model diagnosis module further performs the following steps: The function scheduling submodule triggers the welding defect recognition model, database or external module call based on the judgment of the large language model or user instructions.

6. The process parameter optimization diagnosis method integrating a large language model and a structured model according to claim 4 is characterized in that: The domain language model reasoning submodule is a large language model that is pre-trained with welding domain corpus and fine-tuned with instructions based on Qwen2.5-1.5B.

7. The process parameter optimization diagnosis method integrating a large language model and a structured model according to claim 1 is characterized in that: The interactive interface and intelligent feedback display module performs the following steps: The user input parsing submodule analyzes the queries, instructions, and conversation content raised by the user; The multi-round interaction tracking submodule records and processes multi-round interaction information in a continuous context; The visual display submodule displays the diagnosis results to the user in the form of charts, text or mixed formats; The user feedback collection submodule collects user correction opinions and annotation feedback information; The permission control and deployment interface submodule manages user access rights and supports local / remote deployment of the system.

8. A process parameter optimization and diagnosis system integrating a large language model and a structured model, characterized by including: Process data acquisition and preprocessing module, used to obtain and process process parameter data during welding; A welding defect recognition model building module is used to identify and predict welding defects using a welding defect recognition model; The structured output alignment and language model diagnosis module is used to align structured prediction results to natural language prompts and perform diagnostic reasoning through language models; The interactive interface and intelligent feedback display module are used to display the diagnosis results and collect user feedback information.

Images