Industrial quality data management system based on dialogue type large language model

Abstract

The invention discloses an industrial quality data management system based on a dialogue type large language model, which belongs to the technical field of artificial intelligence and industrial quality management, and comprises the following steps: constructing a quality data management function module comprising statistical analysis, graph generation and data processing, and establishing an intelligent instruction analysis module and adopting a mixed retrieval technology to obtain a quality data management result; high-precision conversion from a natural language instruction of a user to a standardized task instruction is realized, so that a non-professional user can directly complete a professional quality data analysis task through a dialogue mode without mastering complex statistical knowledge; by constructing a semantic understanding model based on a large language model fine tuning technology, accurate understanding of professional terms and complex instructions in the field of industrial quality data management is realized; by setting the semantic slot type contained in the intention enhancement component prediction instruction, the accuracy of user intention understanding is improved. The whole technical scheme realizes a complete intelligent closed loop from user natural language input to task execution.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and industrial quality management, and particularly relates to an industrial quality data management system based on a dialog large language model. BACKGROUND

[0002] Traditional industrial quality data management mainly relies on manual demand analysis and data extraction, and obtains analysis conclusions through statistical analysis tools. This method is not only inefficient, but also requires users to have professional statistical knowledge, and the operation process is complex and complicated.

[0003] In recent years, deep learning technology has been applied in industrial quality data management, and specific quality management models are trained to replace part of the manual operation. However, these methods are often designed for specific tasks and are difficult to adapt to diversified quality data management needs. When enterprises need to handle multiple quality data management tasks at different stages of production, users need to frequently switch between different statistical analysis tools or platforms, which is complex and prone to errors.

[0004] With the rise of large language model technology, natural language processing capabilities have been significantly improved, providing a new technical path to solve the above problems. By building a dialog system based on a large language model, users can directly express quality data management requirements using natural language, and the system automatically understands user intent and performs corresponding tasks, thereby greatly improving work efficiency. SUMMARY

[0005] In view of the above problems, the present application is proposed.

[0006] Therefore, the technical problem solved by the present application is: how to build an intelligent quality data management system that can understand natural language instructions, realize complex quality data analysis tasks through dialog mode, avoid the problems of complex operation, low efficiency and the need for professional knowledge in traditional methods, and provide a one-stop intelligent quality data management solution for enterprises.

[0007] To solve the above technical problems, the present application provides the following technical solutions: an industrial quality data management system based on a dialog large language model, comprising, including, constructing a quality data management function module, including statistical analysis, graph generation and data processing; establishing an intelligent instruction analysis module, converting user input instructions into standardized task instructions through hybrid retrieval technology; constructing a semantic understanding model, based on large language model fine-tuning technology to analyze user intent and parameter information in the standardized task instruction; setting an intent enhancement component to predict the semantic slot type contained in the standardized task instruction to improve understanding accuracy; According to the parsed user intention and parameter information, a corresponding quality data management function module is matched and called to perform a task.

[0008] As a preferred scheme of the industrial quality data management system based on the dialogue type large language model, the intelligent instruction analysis module comprises: a standard instruction template database is constructed, and the standard instruction template database contains standard instruction templates of quality data management tasks; a mixed retrieval technology is used to obtain candidate templates related to the user instruction; an optimal instruction template is selected by using a reordering model; and the user instruction is rewritten into a standard instruction based on the optimal template.

[0009] The beneficial effects of the preferred technical scheme are: when an industrial quality control engineer inputs a colloquial instruction "help me check whether the data in column C2 conforms to the normal distribution", through the three-level processing mechanism of matching the standard instruction template database, screening the candidate templates by the mixed retrieval technology, and selecting the optimal template by the reordering model, the instruction is automatically converted into a standard instruction "perform normality test on column C2, the percentile line is 95, and the data value is located" that can be recognized by the system, so that the complex task that needs to master statistical professional terms and software operation specifications is transformed into a simple dialogue interaction that can be completed by the engineer using natural language, and the technical threshold requirement of the traditional quality data management system for the user's professional background is eliminated.

[0010] As a preferred scheme of the industrial quality data management system based on the dialogue type large language model, the mixed retrieval technology comprises: The document retriever and the vector retriever are used to retrieve candidate templates in parallel; the two retrieval results are integrated by using a fusion algorithm; and the fusion result is refined by using a reordering model to obtain an optimal instruction template.

[0011] As a preferred scheme of the industrial quality data management system based on the dialogue type large language model, the construction of the semantic understanding model comprises: The open source large language model is used as a basic framework; a field-specific training data set is constructed, and the field-specific training data set contains instructions and corresponding structured annotations; the instruction fine-tuning is performed on the basic model by using the fine-tuning technology; and the accurate analysis of the standard instruction is realized by using the intention recognition and slot filling.

[0012] The beneficial effects of the preferred technical solution are: for the highly specialized terms of "Anderson-Darling test", "process capability index Cpk", and the repeatability and reproducibility of measuring instruments in the field of industrial quality data management, by constructing a domain-specific training dataset containing instruction-structured annotation correspondence on the basis of the Qwen2-1.5B-Instruct base model, and using DoRA fine-tuning technology to fine-tune the instructions, the model can accurately understand these professional terms and their parameter requirements which rarely appear in general corpus, solve the understanding deviation and parameter extraction error problem caused by the scarcity of training corpus when general large language models face professional vocabulary in industrial vertical fields, and ensure accurate parsing and execution of complex quality management instructions.

[0013] As a preferred scheme of the industrial quality data management system based on the dialog large language model, the fine-tuning technology comprises: Inserting parameter efficient structure in the attention module and multi-layer perception of the model; decomposing the weight parameter into a direction vector and an amplitude vector; Further decomposing the direction matrix using a low-rank adaptive method; only updating the decomposed low-rank matrix parameters to improve training efficiency.

[0014] As a preferred scheme of the industrial quality data management system based on the dialog large language model, the intent enhancement component comprises: Using a text encoder to obtain a vector representation of the instruction; extracting local features through a convolutional neural network; predicting the slot types contained in the instruction using a fully connected layer; inputting the predicted slot type information as a prompt into the semantic understanding model.

[0015] As a preferred scheme of the industrial quality data management system based on the dialog large language model, the quality data management function module comprises: A random data generation module for generating sample data of various probability distributions; a graph generation module for creating statistical charts and visual results; a statistical analysis module for performing statistical tasks of hypothesis testing, regression analysis, and process control.

[0016] As a preferred scheme of the industrial quality data management system based on the dialog large language model, the function module calling and matching according to the parsed user intent and parameter information comprises: determining the target execution module according to the recognized function name; converting the parsed parameter information into the input format required by the module; calling the corresponding module to execute the quality data management task; returning the execution result to the user in the form of a chart or a numerical value.

[0017] The beneficial effects of the present application: the present application constructs an industrial quality data management system based on a dialogue type large language model. When a quality control engineer faces a complex task of statistical process control analysis on multiple batches of product data, the traditional way requires opening professional statistical software such as Minitab one by one, manually importing each batch of data file, selecting control chart type (X-bar chart, R chart, S chart, etc.) one by one, setting control limit parameters, generating charts and manually judging process stability. The whole process needs to frequently switch between multiple software interfaces and requires the operator to have a deep statistical background. The present application converts the user's natural language instruction "analyze the process stability of these batches of data and generate control charts" into a standardized task sequence executable by the system through an intelligent instruction analysis module. The semantic understanding model accurately identifies the user's intention and parameter requirements, directly calls the corresponding statistical analysis function module to complete complex mathematical calculations and chart generation, thereby simplifying the multi-step professional operation that requires a senior quality engineer to spend 30-45 minutes and is prone to errors into an intelligent process that can be completed by ordinary production management personnel in 12 seconds with only one sentence, eliminating the dependence of industrial quality data management on professional statistical software operation skills, and realizing the popularization and efficiency of quality management work. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 The overall flowchart of an industrial quality data management system based on a dialogue type large language model provided by an embodiment of the present application.

[0020] Figure 2 The graphical summary output result schematic diagram of an industrial quality data management system based on a dialogue type large language model provided by an embodiment of the present application.

[0021] Figure 3 The instruction alignment module diagram of an industrial quality data management system based on a dialogue type large language model provided by an embodiment of the present application.

[0022] Figure 4 The instruction fine-tuning module diagram of an industrial quality data management system based on a dialogue type large language model provided by an embodiment of the present application.

[0023] Figure 5A slot type prediction module diagram of an industrial quality data management system based on a conversational large language model is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0024] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0025] Embodiment 1, Reference Figures 1-5 For an embodiment of the present application, the embodiment provides an industrial quality data management system based on a conversational large language model, comprising: S100, constructing a quality data management function module, including statistical analysis, graph generation and data processing; S200: establishing an intelligent instruction analysis module, converting user input instructions into standardized task instructions through hybrid retrieval technology; S300: constructing a semantic understanding model, analyzing user intent and parameter information in the standardized task instruction based on large language model fine-tuning technology; S400: setting an intent enhancement component, predicting the semantic slot type contained in the standardized task instruction to improve understanding accuracy; S500: according to the analyzed user intent and parameter information, matching and calling the corresponding quality data management function module to execute the task.

[0026] The current industry is transforming towards flexible and diverse production modes, and the complexity of industrial quality data management is also increasing. Many industrial enterprises mainly rely on traditional statistical analysis systems in quality management. These systems usually have two obvious shortcomings: first, they cannot effectively integrate and manage quality-related data, leading to data island phenomenon, and thus cannot realize comprehensive and continuous tracking and monitoring of product quality; second, enterprises often need to handle multiple quality data management tasks at different stages of production, and need to frequently switch between different statistical analysis modules or even different platforms in the use process, which is cumbersome and prone to errors, and low in efficiency.

[0027] To solve the above problems, the quality data management function module constructed by S100 provides a one-stop quality data management platform for users by integrating statistical analysis, graph generation, and data processing functions, avoiding the need to switch between multiple tools in the traditional way; the intelligent instruction analysis module established by S200 uses the Retrieval Augmentation Generation (RAG) technology, which can convert the user's natural language instructions into standardized task instructions that the system can understand, solving the problem that the traditional system requires users to have professional statistical knowledge; the semantic understanding model constructed by S300 is based on large language model fine-tuning technology and is optimized specifically for the field of industrial quality data management to ensure accurate understanding of domain-specific terminology and complex instructions; the intent enhancement component set by S400 further improves the accuracy of the system's understanding of user intent by predicting the semantic slot types contained in the instructions; S500 intelligently matches and calls the corresponding function modules according to the analyzed user intent and parameter information, realizing a complete closed loop from user instructions to task execution.

[0028] The English characters involved in the present application are explained as follows: Anderson-Darling test (Anderson-Darling-Test, AD for short) is a statistical method for testing whether data conforms to a certain distribution; process capability index Cpk (Process-Capability-Index, Cpk for short) is a statistical index for measuring the ability of a process to meet specification requirements; Qwen2-1.5B-Instruct (Qwen2-1.5B-Instruct) is a 1.5 billion parameter instruction fine-tuning version of the 100 million parameter instruction of the open source large language model; DoRA (Dynamic-Rank-Adaptation, DoRA for short) is an efficient fine-tuning technology for dynamic rank adaptation; Minitab (Minitab) is a professional statistical analysis software; m3e-base model (Multilingual-E5-base, m3e-base for short) is a multilingual text embedding model; BM25 retriever (Best-Matching-25, BM25 for short) is an information retrieval algorithm based on a probability ranking function; KNN retriever (K-Nearest-Neighbors, KNN for short) is a K-nearest neighbor retrieval algorithm; Large model Qwen-Max (Qwen-Max) is a large-scale language model of the Qwen series; LoRA (Low-Rank-Adaptation, LoRA for short) is a low-rank self-adaptive fine-tuning method; AdamW optimizer (Adaptive-Moment-Estimation-with-Weight-decay, AdamW for short) is an adaptive moment estimation optimization algorithm with weight decay; token is the smallest semantic unit in text processing; prompt is a prompt word input to a large language model; ALBERT (A-Lite-BERT, ALBERT) is a lightweight bidirectional encoder representation model; BERT model (Bidirectional-Encoder-Representations-from-Transformers, BERT for short) is a bidirectional encoder representation model based on Transformer; RoBERTa (Robustly-Optimized-BERT-Pretraining-Approach, RoBERTa for short) is a robustly optimized BERT pre-training method; DeBERTa (Decoding-enhanced-BERT-with-Disentangled-Attention, DeBERTa for short) is a decoding-enhanced BERT model with disentangled attention; Dropout is a regularization technique to prevent overfitting.n-gram is a combination feature of continuous n words; SluQwen is a large model (Spoken-Language-Understanding-Qwen, English abbreviation: SluQwen) specially used for spoken language understanding; WebSocket is a network protocol for bidirectional communication; AGIF is an attention-based graph information fusion model (Attention-based-Graph-Information-Fusion, English abbreviation: AGIF); GL-GIN is a graph learning model based on graph isomorphism network (Graph-Learning-with-Graph-Isomorphism-Network, English abbreviation: GL-GIN); SDJN is a slot-gated dual joint network (Slot-gated-Dual-Joint-Network, English abbreviation: SDJN); GISC is a graph interactive semantic classifier (Graph-Interactive-Semantic-Classifier, English abbreviation: GISC); Co-guiding-Net is a co-guiding network; ReLa-Net is a relation-aware network (Relation-aware-Network, English abbreviation: ReLa-Net); SSRAN is a slot semantic relational attention network (Slot-Semantic-Relational-Attention-Network, English abbreviation: SSRAN); Stephens is a Stephens statistical method (Stephens-Method, English abbreviation: Stephens).

[0029] Embodiment 2, which is an embodiment of the present application, provides an industrial quality data management system based on a conversational large language model based on the previous embodiment, comprising: a quality data management function module is constructed in step S100, step S100 includes the following steps A1-A3: A1: constructing a statistical analysis function module for performing various statistical analysis tasks; Specifically, the statistical analysis function module in step A1 mainly includes 70 kinds of statistical analysis functions such as graphical summary function, experimental design, process capability analysis, regression analysis, hypothesis testing and gauge study. Taking the graphical summary function as an example, this function summarizes numerical data through various statistical quantities (such as sample size, mean value, median, standard deviation), and generates visual images for intuitive analysis by quality control engineers. When the user inputs the table data or transmits the data table in xlsx format into the system, the system first calculates the mean, variance, standard deviation, minimum value, maximum value, median and other basic statistical indicators of the sample, and then performs normality test using the Anderson-Darling method. The specific calculation method is realized through the following formula: {A}^{2}=-n-\frac {1} {n}\sum ^{n}_{i=1} \left [ {\left ( {2i-1} \right )ln{F}_{0}{X}_\left ( {i} \right )+\left ( {2n+1-2i} \right )ln\left ( {1-{F}_{0}{X}_\left ( {i} \right )} \right )} \right ] , Wherein, is the Anderson-Darling test statistic; is the sample size; is the sample index, indicating the ith observation, with a value range of 1 to n; is the natural logarithm function; is the cumulative distribution function of the theoretical distribution, that is, the target distribution from which the sample is assumed to come from; is the ith order statistic, that is, the ith value after sorting the sample from small to large.

[0030] For a given sample , the empirical distribution function is the proportion of observations in the sample that are less than or equal to , which is calculated as follows:

[0031] Wherein, is the empirical distribution function; is the sample size; is the ith observation; is the independent variable of the distribution function, indicating a certain specific numerical value; is the indicator function, equal to 1 when the brackets are true, otherwise equal to 0.

[0032] Theoretical distribution function is the cumulative distribution function of the target distribution assumed to be the source of the sample, calculated as follows: , where, is the theoretical distribution function; is the exponential function; is the integral variable, representing the continuous values from negative infinity to x during the integration process; and are the sample mean and sample standard deviation, respectively; is the differential element, representing the integration of variable t.

[0033] Finally, the P value of the Anderson-Darling normality test is calculated using the Stephens method. First, the original statistic is modified to conform to a nearly stable distribution: , where, is the modified Anderson-Darling statistic, used to conform to a more stable distribution; is the original Anderson-Darling statistic; is the sample size.

[0034] Then, the p value is calculated using the Stephens piecewise approximation formula:

[0035] , where, is the piecewise function, used to handle input values in different ranges; is the function input value, here referring to the modified statistic ; is the modified Anderson-Darling statistic, used as the judgment condition and calculation input of the piecewise function; is.

[0036] It should be noted that the design purpose of the statistical analysis function module is to provide comprehensive statistical analysis capabilities for industrial quality data management. Through modular architecture design, each function can be independently called, while supporting batch processing and real-time analysis. The module uses a unified data interface to ensure seamless integration with other modules.

[0037] As Figure 2 shown, it is a graphical summary output result schematic diagram of the present application. When the user inputs the data to be analyzed, the system automatically calculates the basic statistical indicators and performs normality test using the Anderson-Darling method, generating a visual result containing a histogram and a normal distribution curve.Figure 2 The complete analysis report of system output is displayed, including sample statistics summary table, normality test result, and histogram and fitted normal distribution curve of data distribution. The normality test result is Anderson-Darling statistic and P value, and the sample statistics summary table is a table showing sample number, mean, standard deviation, minimum value, maximum value, etc. The user can intuitively judge the normality of data according to the overlapping degree of the histogram and the normal distribution curve, so as to quickly identify data anomalies.

[0038] A2: Construct a graphics generation function module for creating statistical charts and visual results; Specifically, the graphics generation function module in step A2 includes scatter plot, bubble chart, time series chart, histogram, box plot, Pareto chart, and other 11 kinds of statistical image generation functions.

[0039] A3: Construct a data processing function module for generating and processing various data types.

[0040] In the present embodiment, the intelligent instruction analysis module in step S200 includes the following steps B1-B4: B1: Construct a standard instruction template database containing standard instruction templates for quality data management tasks; In the quality data management scenario, a team of experts in the field is organized, and based on the enterprise's internal quality management process and historical task records, a template data set containing 362 standard task instructions is manually written, including instructions corresponding to all function modules. For each specific function module, there are usually 1 to 5 template variants. For example, for the "normality test" function, the instruction template is set to "test the normality of <value1>Column execution normality test, percentile line is <value2>, located at data value " and " pair <value1>Column execution normality test, percentile line is <value2>, located at Y value ", corresponding to two kinds of percentile line options under different templates. Among them <valuen>For the content to be filled in the template, fill in this part according to the original instruction in the subsequent large model rewriting step. Then, based on the data set, an instruction template document database and an instruction template vector database are respectively constructed to realize multi-channel mixed retrieval.

[0041] B2: Obtain candidate templates related to the user instruction through the mixed retrieval technology; Use the document retriever and the vector retriever to retrieve candidate templates in parallel; Integrate the two retrieval results through a fusion algorithm; Use a reordering model to fine-tune the fusion results to obtain the optimal instruction template.

[0042] Specifically, when a new function is put online or a new quality management scenario appears, the quality data management field expert team collaborates to write the corresponding new instruction template and synchronously updates the two databases; at the same time, the retrieval log and recall effect are analyzed regularly, and inefficient or abandoned templates are marked for revision or deletion, and the index is synchronously updated; in the document retrieval channel, the document is first cut into blocks according to the sentence, and each block corresponds to a different instruction template. In the vector retrieval channel, all task instruction templates are converted into 768-dimensional sentence vectors through the m3e-base model. When the user inputs the instruction, the BM25 retriever retrieves m instruction templates related to the quality data management field terms in the user instruction, and the weighting mechanism can better match the importance of the keywords in the vertical field; the KNN retriever retrieves in the vector database to maximize the semantic relevance of the instruction templates and the user instruction, and obtains the most similar k instruction templates. Among them, m and k are adjustable hyperparameters, which are both set to 5 by default, and 10 relevant instruction templates are finally obtained through retrieval. Since the word embedding model encodes the original instruction and the instruction template into a fixed-dimensional vector respectively, such as the m3e-base word vector dimension of 768 used in this scheme, the rich semantics of the text is compressed into a low-dimensional vector, which cannot completely preserve all the potential meanings of the original text, so the results of the retrieval often still contain some instruction templates with poor relevance.

[0043] It should be noted that the design purpose of the mixed retrieval technology is to make full use of the advantages of different retrieval methods, the BM25 retriever is good at accurate matching of professional terms, and the KNN retriever is good at semantic similarity matching, and the combination of the two can improve the accuracy and recall rate of retrieval. The system triggers two retrievals simultaneously in an asynchronous parallel calling manner, and the overall retrieval delay is stable within 100ms.

[0044] B3: Use the reordering model to select the optimal instruction template; Specifically, the reordering model in step B3 uses bge-reranker-base to fine-rank the 10 instruction templates retrieved. Bge-reranker-base is a text matching model of cross-encoder architecture, which can calculate a fine-grained similarity score between an original instruction and an instruction template. Its core workflow is as follows: first, the cross-encoder concatenates the "user original instruction" and the "instruction template" into a complete input sequence to generate a single embedding, thereby capturing the representation and correlation of the sentence, and then sends it to the feedforward neural network to output a scalar correlation score between 0 and 1, with a higher score indicating stronger correlation. For the 10 instruction templates retrieved, the correlation score between each and the original instruction is calculated according to the above process, and the one with the highest correlation is finally retained as the optimal instruction template.

[0045] B4: Rewrite the user instruction into a standard instruction based on the optimal template.

[0046] Specifically, after obtaining the optimal instruction template in step B4, it is input into the large model Qwen-Max together with the user input task instruction, and the prompt for rewriting is: "Please rewrite the following task instruction {Q} in spoken language according to the template {T}, and the <valuen>Placeholders represent replaceable text or values. Where T and Q are placeholders that need to be replaced with specific templates and instructions. Ultimately, the large model translates the user instructions into appropriate standard instructions based on this command. The instruction alignment module graph generated based on the retrieval enhancement is shown in FIG. 8, ensuring that the output format meets the input requirements of the subsequent semantic understanding model. During the rewriting process, the system maintains the key information in the original instructions while standardizing the expression and eliminating colloquial expressions and non-standard words. Figure 3

[0047] In this embodiment, the construction of the semantic understanding model in step S300 includes the following steps C1-C3: C1: Based on an open-source large language model as the basic architecture; Specifically, in step C1, the open-source large language model Qwen2 is selected as the basic architecture. Qwen2 is a large language model using the transformer decoder architecture, pre-trained on a large-scale dataset of over 18T tokens, with strong natural language processing capabilities, capable of completing instruction execution, text generation, and structured output tasks. The model structure of Qwen2 mainly includes a text input layer, an embedding layer, a decoder layer, and an output layer. The system selects the smaller Qwen2-1.5B-Instruct model in the Qwen2 series as the basic model, which has lower computational resource requirements while maintaining good performance, suitable for deployment in industrial environments. The selection of the model considers the balance of deployment cost, inference speed, and accuracy, and Qwen2-1.5B-Instruct exhibits good understanding ability and generalization performance on quality data management tasks.

[0048] It should be noted that the selection of the basic model is a key decision in the construction of the semantic understanding system, which needs to consider multiple factors such as the language understanding ability of the model, the consumption of computational resources, the convenience of deployment, etc. The open-source nature of the Qwen2 model ensures the controllability and security of the system, avoiding dependence on external API services.

[0049] C2: Construct a domain-specific training dataset containing instructions and corresponding structured annotations; Specifically, in step C2, based on the standard instruction template dataset constructed manually in the instruction alignment module, random values are used to replace the placeholders in the instruction templates, and the corresponding structured annotations are added to form the domain-specific training dataset. The structured annotations include the instruction type, the instruction execution target, and the instruction execution result, etc. The structured annotations are used to guide the training of the semantic understanding model, ensuring that the model can accurately understand the user's instructions and generate the corresponding structured output. <valuen>The templates are replaced and augmented using a large language model to construct a fine-tuning dataset. Taking the template of the "normality test" function as an example, the original template "to <value1>Column execution normality test, percentile line is <value2>, located at data value " by generating a random number to convert to "Perform normality test on column C2, test type is Anderson-Darling, reference line type is 52, located at data value ", and construct the corresponding intent and semantic slot structured data, the code is as follows: { "functionName" : "normalityTest", "parameters" : { "columnIndex" : 2, "referenceLineType" : 2, "values" : [ 52 ], "testType" : 0 } } Thus, the annotation sample <text, groundtruth> is constructed: { "text" : "Perform normality test on column C2, test type is Anderson-Darling, reference line type is 52, located at data value " "groundtruth" : { "functionName" : "normalityTest", "parameters" : { "columnIndex" : 2, "referenceLineType" : 2, "values" : [ 52 ], "testType" : 0 } } } Apply this method to all templates to generate 20000 formatted samples. Since the templated samples often have single sentence patterns, the application further uses a large model to perform data augmentation on the samples, thereby enriching the data set.

[0050] The prompt word used is: "You are a data augmentation assistant. Please rewrite the following instruction in a new expression, requirements: 1. The style can be formal, colloquial, or scenario-based description; 2. Must retain the core elements in the sentence;" The formatted sample "Perform normality test on C2 column, percentile line is 52, located at data value" and the prompt are input into the Owen-max large model to obtain the enhanced instruction: "Can you help me perform the Anderson-Darling normality test on the data? My sample is in column C2, and the percentile line is marked at data value 52" The enhanced instruction is added to the data set, and finally a data set containing 50,000 diversified instructions is obtained, which is divided into training set, validation set and test set in the ratio of 8:1:1.

[0051] In the fine-tuning stage, first insert additional DoRA structure into the attention module and multilayer perceptron of each decoder layer of the model, which contains two independent parts of amplitude vector m and direction vector V uniformly decomposed from pre-training weight parameters. Then, using the low-rank adaptive fine-tuning method LoRA, further decompose the direction matrix V into low-rank matrices A and B, while the amplitude part remains unchanged, and freeze the weight parameters of the large model itself. Finally, use the AdamW optimizer to update only matrices A and B. In training, the autoregressive cross-entropy loss based on language modeling objective is used, and for each piece of text, only the cross-entropy loss of the tokens predicted by the model for the target part is calculated, and the padding and input prompt part are ignored. mask. Since two low-rank matrices A and B are used to approximately represent ΔV, where the rank r is much smaller than d, the original The parameter quantity is reduced to without changing the dimension of the output data, effectively reducing resource consumption in training. In training, the batch size is set to 16, the rank r of LoRA is set to 32, the training rounds are set to 5, and the learning rate is set to 0.00005. After fine-tuning, a spoken language understanding model specialized for the field of industrial quality data management is obtained. The module diagram is shown in the accompanying Figure 4 .

[0052] C3: Fine-tune the base model using fine-tuning technology; Insert parameter-efficient structure into the attention module and multilayer perceptron of the model; decompose the weight parameters into direction vectors and amplitude vectors; Further decompose the direction matrix using a low-rank adaptive method; only update the low-rank matrix parameters after decomposition to improve training efficiency.

[0053] Specifically, in step C3, the base model is fine-tuned using the DoRA (Dynamic Rank Adaptation) fine-tuning technique. In the fine-tuning phase, first, additional DoRA structures are inserted into the attention modules and multi-layer perceptrons of each decoder layer of the model, which include two independent parts, an amplitude vector m and a direction vector V, which are uniformly decomposed from the pre-training weight parameters. Then, the low-rank adaptive fine-tuning method LoRA is used to further decompose the direction matrix V into low-rank matrices A and B, while the amplitude part remains unchanged, and the weight parameters of the large model itself are frozen. Finally, the AdamW optimizer is used to update only matrices A and B. In training, the self-recurrent cross-entropy loss based on the language modeling objective is used, and only the cross-entropy loss of the tokens predicted by the model for the target part of each text is calculated, and the padding and input prompt parts are ignored. Since two low-rank matrices A and B are used to approximately represent ΔV, where the rank r is much smaller than d, the original parameter amount of dxd is reduced to 2xrxd without changing the dimension of the output data, effectively reducing the resource consumption in training.

[0054] It should be noted that the core advantage of the DoRA fine-tuning technique is parameter efficiency and training stability. By decomposing the pre-training weight and updating only the low-rank part, the computational cost and storage requirements of fine-tuning are significantly reduced. In training, the batch size is set to 16, the rank r of LoRA is set to 32, the training round is set to 5, and the learning rate is set to 0.00005.

[0055] C4: Precise analysis of standard instructions through intent recognition and slot filling.

[0056] Specifically, in step C4, precise analysis of standard instructions is achieved through intent recognition and slot filling. Intent recognition is responsible for determining the function module corresponding to the user instruction, such as "normality test" and "regression analysis"; slot filling is responsible for extracting specific parameter information in the instruction, such as column index, numerical value, and options.

[0057] In the present embodiment, the intent enhancement component in step S400 includes the following steps D1-D4: D1: Obtain a vector representation of the instruction using a text encoder; Specifically, ALBERT is used as the text encoder in step D1 to obtain word vectors. ALBERT (ALiteBERT) is a parameter-shared lightweight pre-training language model with fewer parameters and faster inference speed than the BERT model, while maintaining good text representation capabilities. In this system, the ALBERT encoder converts the input instruction text into a fixed-dimensional vector representation, with each token corresponding to a 768-dimensional vector. The encoding process includes the combination of word embedding, position embedding, and segment embedding, which are processed through multiple layers of transformer encoders to obtain text representations containing rich semantic information. The parameter sharing mechanism of ALBERT significantly reduces the parameter quantity while maintaining the expression ability, making it suitable for deployment in resource-constrained industrial environments.

[0058] It should be noted that the selection of the text encoder considers the balance between encoding quality, computational efficiency, and deployment convenience. The ALBERT model performs well in understanding professional terms in the quality data management field and can effectively capture key information in the instructions.

[0059] D2: Extract local features through a convolutional neural network; Specifically, in step D2, single-layer convolution and max-pooling are performed through a text convolutional neural network to further extract local features. Considering that domain-related task instruction texts are often long and have complex semantic relationships, and the length of domain-specific vocabulary contained in them varies, the system sets convolution kernels with heights of 2, 3, 4, 5, 6, and 7 to capture local features of different ranges. Finally, the outputs of multiple convolution kernels are concatenated to form comprehensive local feature representations.

[0060] It should be noted that the design of the convolutional neural network fully considers the language characteristics of quality data management instructions, capturing professional terms and phrase patterns of different lengths through multi-scale convolution kernels, improving the effectiveness of feature extraction.

[0061] D3: Predict the slot types contained in the instructions using a fully connected layer; Specifically, in step D3, the concatenated local features are projected into the predicted slot types using a fully connected layer. The fully connected layer uses a multi-layer perceptron structure, containing a hidden layer and an output layer. The hidden layer uses a ReLU activation function, and the output layer uses a sigmoid activation function to support multi-label classification, as an instruction may contain multiple different slot types. The goal of slot type prediction is to identify the parameter types that may be contained in the instruction, such as columnIndex (column index), title (title), referenceLineType (reference line type), values (numeric value), testType (test type), etc.

[0062] Finally, multiple local features are concatenated and projected into the predicted slot type using a fully connected layer. The instruction and slot type are extracted from the fine-tuning dataset as the slot type prediction module dataset, and the slot type is the key of the parameters field, for example: { "text" : "Perform normality test on C2 column, test type is Anderson-Darling, percentile line is 52, located in data value" "slotType" : ["columnIndex", "title", "referenceLineType", "values", "testType"] } And also divided into training set, validation set and test set in the ratio of 8:1:1. When training the slot type prediction module, cross entropy is used as the loss function, AdamW optimizer is selected, the initial learning rate is set to 3x10 -5 , the weight decay coefficient is set to 0.01; the batch size is 32, the training round is set to 20, and the early stopping mechanism is triggered when the validation set loss does not decrease for two consecutive rounds. In order to prevent overfitting, 0.2 Dropout is added after the convolutional layer and the fully connected layer, and the learning rate warm-up strategy is adopted, and the learning rate is linearly increased from 0 to 3x10 -5 in the first 250 gradient updates, which accounts for about 10% of the total training steps, and then linearly decays to 0 in the remaining training steps to obtain the best generalization performance. Each convolution kernel performs sliding convolution operation on the text vector sequence to extract n-gram level feature patterns. After convolution operation, the maximum pooling layer is used to select the maximum value in each feature map to retain the most important feature information. Convolution kernels of different heights can capture semantic features of different granularities from double-word combinations to seven-word combinations, which is very important for understanding complex professional terms in quality data management.

[0063] In the embodiment of the application, the full connection layer in step D3 can also introduce hierarchical slot type classification, group the slot types according to the function modules, and use a hierarchical classifier for prediction.

[0064] D4: Input the predicted slot type information into the semantic understanding model as a prompt.

[0065] The module structure is shown in the accompanying Figure 5 Before large model inference, the slot type prediction module is used to obtain possible slot types and fill them as prompt words: The semantic slots that may be contained in the instruction are: "columnIndex", "title", "referenceLineType", "values", "estType". Splicing it with the instruction gets "Perform normality test on column C2, test type is Anderson-Darling, reference line type is percentile, and values are data values. The semantic slots that may be contained in the instruction are: "columnIndex", "title", "referenceLineType", "values", "testType". Input the SluQwen large model, thereby enhancing the accuracy of the large model's intent understanding.

[0066] In the experiment, the overall accuracy of the model on the test set reached 95.23%, that is, in more than 95% of the samples, the module can correctly predict the slot type contained in the user instruction, which can effectively assist the large model to improve the intent understanding ability.

[0067] In this embodiment, the function module is matched and called according to the parsed user intent and parameter information in step S500, including the following steps E1-E7: E1: Determine the target execution module according to the recognized function name; Specifically, in step E1, the target execution module is determined according to the functionName field output by the semantic understanding model. The system maintains a function module mapping table, which maps each function name to a specific execution module. For example, "normalityTest" corresponds to the normality test module, "regressionAnalysis" corresponds to the regression analysis module, and "processControl" corresponds to the process control module. The mapping table uses a hash table structure, supporting fast lookup with O(1) time complexity. When receiving a function name, the system first verifies whether the function exists in the mapping table, and if it exists, obtains the corresponding module reference, and if it does not exist, returns an error message prompting the user. The module mapping table supports dynamic updating, and when a new function module is added, it can be registered through a configuration file or a management interface to ensure the scalability of the system. Each function module implements a unified interface specification, including input parameter verification, task execution, result formatting, and other standard methods.

[0068] It should be noted that the organization and management of the function module adopts a plug-in architecture design, and new analysis functions can be developed and deployed as independent modules without affecting the stability of the existing system.

[0069] In the embodiment of the present application, the target execution module in step E1 is determined according to the functionName field output by the semantic understanding model, and the system maintains a function module mapping table to correspond each function name to a specific execution module.

[0070] E2: converting the parsed parameter information into an input format required by the module; Specifically, in step E2, the parameter information in the parameters field output by the semantic understanding model is converted into an input format required by the target module. Different function modules have different requirements for input parameters, and the system needs to perform format conversion and type checking. Taking the normality test module as an example, the input parameters required by the module include data column index (integer type), test type (enumeration type), confidence level (floating point type), reference line setting (dictionary type), etc. The system first performs type verification on each parameter to ensure that the numerical parameters are within a reasonable range, the enumeration parameters are in the allowed value set, and the string parameters meet the format requirements. Then, unit conversion and standardization processing are performed, such as converting percentages to decimals, converting text descriptions to numerical codes, etc. For missing optional parameters, the system will use preset default values to fill in. If incompatible types or out-of-range values are found during parameter conversion, the system will generate detailed error information and return it to the user.

[0071] It should be noted that the design of the parameter conversion module fully considers the diversity and fault tolerance of user input, and through intelligent type inference and format conversion, the usability of the system is maximized.

[0072] In the embodiment of the present application, the parsed parameter information in step E2 is the parameter information in the parameters field output by the semantic understanding model, and the parameter information in the parameters field output by the semantic understanding model is converted into an input format required by the target module.

[0073] E3: calling the corresponding module to execute the quality data management task; Specifically, the process of calling corresponding modules to execute quality data management tasks in step E3 adopts an asynchronous execution framework. For computationally intensive tasks such as large data volume statistical analysis, the system submits the task to the background thread pool for execution, avoiding blocking the user interface. During task execution, the system monitors the execution status and progress in real time and pushes progress updates to the front end through the WebSocket connection. Each functional module performs exception handling and resource management during execution to ensure that the overall stability of the system is not affected even in the event of errors. For long-running tasks, the system supports task cancellation and suspension functions, and users can interrupt unnecessary calculations at any time. After task execution is complete, the system stores the results in a temporary cache and generates a unique task ID for user query and download. At the same time, the system records detailed execution logs, including input parameters, execution time, resource consumption, and other information, for performance analysis and problem troubleshooting.

[0074] It should be noted that the design of the asynchronous execution framework improves the concurrent processing capability and user experience of the system, supporting multiple users to use simultaneously without affecting each other.

[0075] In the embodiment of the application, the process of calling corresponding modules to execute quality data management tasks in step E3 adopts an asynchronous execution framework. For computationally intensive tasks such as large data volume statistical analysis, the system submits the task to the background thread pool for execution, avoiding blocking the user interface.

[0076] E4: Return the execution results to the user in the form of charts or numerical values.

[0077] Specifically, the process of returning the execution results to the user in the form of charts or numerical values in step E4 includes three main links: result formatting, visualization generation, and transmission optimization. In the result formatting stage, the system generates the corresponding output format according to different analysis types, such as statistical summary table, hypothesis testing report, regression analysis results, etc. In the visualization generation stage, the system automatically selects the most suitable chart type to display the analysis results, such as histogram to display data distribution, scatter plot to display correlation, control chart to display process stability, etc.

[0078] In this embodiment, the quality data management functional module in step S500 includes the following steps E5-E7: E5: Random data generation module for generating sample data of various probability distributions; E6: Graph generation module for creating statistical charts and visual results; E7: Statistical analysis module for performing hypothesis testing, regression analysis, and process control statistical tasks.

[0079] In this embodiment, all front-end and back-end and inter-module network communications are encrypted and transmitted through the TLS 1.2 / 1.3 protocol to ensure the confidentiality and integrity of the data during transmission. The sensitive fields stored in the database are symmetrically encrypted by the AES-256 algorithm when writing, including the original quality data and user identity information. All encryption keys are centrally managed by a dedicated key management service KMS, and key rotation is performed regularly to prevent key leakage or misuse.

[0080] The system adopts a role-based access control scheme, divides user accounts into "administrator", "quality data management engineer", "operation and maintenance personnel" and other different roles, and each role is only authorized to call quality management functions within its scope of responsibility. All key operations, including template library updates of the instruction alignment module, updates of the spoken language understanding large model, and key scheduling, must pass through multi-factor authentication, which requires dynamic passwords or hardware tokens in addition to username / password verification, thereby eliminating the risk of unauthorized access caused by the cracking of a single credential.

[0081] When the user-submitted instruction text and quality data contain enterprise confidential information, the system will first perform simple regular matching and keyword replacement on these data and text. Common sensitive information, such as the actual name, employee number, and part number in the operator field of the quality data, is replaced with uniform placeholders (such as <name>, <employee number>, <part number>). The replaced text is used for log recording and as input for model training and inference, avoiding exposure of original sensitive content throughout the process. For historical dialogue data, a monthly inventory file is exported, and sensitive fields are scanned and replaced in batches through the same script, and then stored to avoid old data leakage. The entire desensitization process is highly scripted and easy to maintain.

[0082] The system generates structured logs for user requests, retrieval calls, model inference, and instruction execution, which record request time, source IP, call interface, input instruction, and execution results and time consumption. Log files are generated on a daily basis and archived to the server storage directory on a weekly basis. The archived logs are compressed and retained for three years to meet subsequent tracing and compliance requirements. The system also automatically sends daily log summary reports to the administrator's email at midnight. The report displays the number of key operations, abnormal failure records, and system errors in table form, allowing administrators to quickly understand the running status.

[0083] In actual operation of the dialogue system, from the user submitting an instruction on the front end to the final return of the interface result, the end-to-end average latency of the system is about 12 seconds - among them, the instruction alignment module takes about 8 seconds due to calling an online large model, SluQwen inference and slot type prediction takes about 1.5 seconds, calling and executing the background function module takes about 1.5 seconds, and the remaining time is used for network transmission and front-end rendering.

[0084] In order to maintain stable performance during peak hours, the present scheme adopts containerized microservice deployment, and 5 parallel computing nodes are deployed. The system uses a load balancing strategy to evenly distribute active sessions to different nodes, and each node can manage about 500 online sessions, ensuring that there is no queuing or blocking when multiple users operate cooperatively. The overhead of large model inference and backend computing is fully considered, and the availability and real-time performance required in industrial sites are achieved.

[0085] In summary, by constructing quality data management function modules including statistical analysis, graph generation, and data processing, one-stop integration of quality data management tasks is achieved, solving the problem of data silos and the need to frequently switch between multiple platforms in traditional methods, thereby avoiding the drawbacks of complicated and error-prone operations; by establishing an intelligent instruction analysis module and using hybrid retrieval technology, high-precision conversion of user natural language instructions to standardized task instructions is achieved, allowing non-professional users to complete professional quality data analysis tasks directly through dialogue without mastering complex statistical knowledge, thereby significantly reducing the system usage threshold; by constructing a semantic understanding model based on large language model fine-tuning technology, accurate understanding of industrial quality data management field professional terms and complex instructions is achieved, and compared to the limitations of traditional deep learning methods that can only target specific tasks, the present invention can adapt to diverse quality data management needs; by setting an intent enhancement component to predict the semantic slot types contained in the instruction, the accuracy of user intent understanding is further improved, effectively solving the problem of inaccurate understanding of traditional dialogue systems in complex industrial scenarios. The entire technical solution realizes a complete intelligent closed loop from user natural language input to task execution, produces unexpected technical effects of converting traditional complex quality data management work into simple dialogue interaction, and significantly improves the efficiency and intelligent level of industrial quality data management.

[0086] Embodiment 3, refer to Figure 1 and Figure 5 An embodiment of the present invention provides an industrial quality data management system based on a dialogue large language model. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0087] Referring to Table 1: The graphical summary of the numerical data is summarized by various statistical quantities, including sample size, mean, median, standard deviation, and visual images are generated for intuitive analysis by quality control engineers. For example, a quality control engineer needs to ensure that the bottle cap of a medicine bottle is tightly closed. The standard torque value for tightening the bottle cap is 5. The engineer randomly collects a sample containing 10 bottles to check whether the torque required to open the cap is normal.

[0088] The opening torque of the product to be evaluated is shown in Table 1. First, the user inputs the table data or imports the data table in xlsx format into the system, specifies the data column C1 to be analyzed, and inputs the confidence level value. In this example, the confidence level is set to 95%, indicating that if 100 samples are randomly selected from the population, about 95 samples in the confidence interval will contain the population parameter.

[0089] Table 1: Randomly sampled bottle cap torque sample:

[0090] In the fine-tuning effect evaluation stage. Experiments are conducted on the test set using the most commonly used evaluation indicators for spoken language understanding, including intent accuracy, semantic slot F1 score, and overall accuracy, which is only counted as one overall correct when the intent and all slots are correct. The large model without fine-tuning lacks domain knowledge in the field of quality data management, and the parsing effect of the text in this field is very poor. The intent accuracy of the Qwen2-1.5B-Instruct model before fine-tuning is 65.22%, the slot F1 score is only 15.73%, and the overall accuracy is as low as 10.32%. After completing the low-rank adaptive fine-tuning based on DoRA, the model's intent accuracy is improved to 95.03%, the slot F1 score is improved to 94.88%, and the overall accuracy is 92.12%, fully demonstrating the effectiveness of the fine-tuning scheme on the constructed data set and the significant enhancement of the understanding ability of the industrial quality data management scene.

[0091] In order to verify the effectiveness of the semantic understanding model of the present application, a comparative experiment with existing mainstream spoken language understanding technology is designed. The comparative experiment is based on the semantic understanding model construction process of step S300 in embodiment 2, especially the domain-specific training data set containing 50,000 diversified instructions constructed in step C2, and the instruction fine-tuning process of the Qwen2-1.5B-Instruct base model using the DoRA fine-tuning technology in step C3. The experiment selects 7 representative spoken language understanding models as the comparison benchmark, including AGIF, GL-GIN, SDJN, GISC, Co-guiding-Net, ReLa-Net and SSRAN, and evaluates the intent understanding accuracy, slot filling accuracy and overall accuracy of the three key indicators on the same test set. Referring to Table 2: Table 2: Comparison experiments with prior art:

[0092] Referring to the comparison experiment results shown in Table 2, it is shown that the technical scheme proposed by the present application is significantly superior to the prior art in three evaluation indexes. The intent understanding accuracy of the present application is 96.17%, which is improved by 4.68 percentage points compared with the best prior art ReLa-Net 91.49%; the slot filling accuracy is 94.88%, which is improved by 9.54 percentage points compared with the best prior art SSRAN 85.34%; and the overall accuracy is 92.12%, which is improved by 8.66 percentage points compared with the best prior art SSRAN 83.46%. This significant performance improvement fully proves the effectiveness of the technical route of constructing a field-specific training data set, using the DoRA fine-tuning technology, and setting the intent enhancement component, especially in processing professional terms and complex instructions in the field of industrial quality data management, which has obvious technical advantages compared with general spoken language understanding models.

[0093] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application. < / valuen> ​< / valuen> < / valuen>

Claims

1. An industrial quality data management system based on a conversational large language model, characterized in that: The quality data management function module includes statistical analysis, graph generation, and data processing. An intelligent instruction analysis module is established to convert user input instructions into standardized task instructions through hybrid retrieval technology. A semantic understanding model is constructed to analyze user intent and parameter information in standardized task instructions based on large language model fine-tuning technology. An intent enhancement component is set up to predict semantic slot types contained in standardized task instructions to improve understanding accuracy. Based on the analyzed user intent and parameter information, the corresponding quality data management function module is matched and called to execute the task.

2. The industrial quality data management system based on a conversational large language model of claim 1, wherein: The intelligent instruction analysis module includes a standard instruction template database containing standard instruction templates for quality data management tasks.

3. The industrial quality data management system based on a conversational large language model according to claim 2, characterized in that: The hybrid retrieval technology includes using document retriever and vector retriever to retrieve candidate templates in parallel.

4. The industrial quality data management system based on a conversational large language model according to claim 3, characterized in that: The construction of the semantic understanding model includes using an open-source large language model as the basic architecture. A domain-specific training dataset is constructed, containing instructions and corresponding structured annotations.

5. The industrial quality data management system based on a conversational large language model of claim 4, wherein: The fine-tuning technology includes inserting parameter-efficient structures in the attention module and multi-layer perceptron of the base model. The weight parameters are decomposed into direction vectors and amplitude vectors to obtain a direction matrix.

6. The industrial quality data management system based on a conversational large language model of claim 5, wherein: A low-rank adaptive method is used to further decompose the direction matrix.

7. The industrial quality data management system based on conversational large language model of claim 6, wherein: The intent enhancement component includes using a text encoder to obtain a vector representation of the instruction. The quality data management function module includes a random data generation module for generating sample data with various probability distributions. The graph generation module is used to create statistical charts and visual results.

8. The industrial quality data management system based on a conversational large language model of claim 7, wherein: The statistical analysis module is used to perform hypothesis testing, regression analysis, and statistical tasks for process control. Based on the analyzed user intent and parameter information, the corresponding module is matched and called to execute the quality data management task. The execution results are returned to the user in the form of charts.

9. The industrial quality data management system based on a conversational large language model of claim 8, wherein: Also includes security protection module, through TLS1.2 / 1.3 protocol to the front and back end and the network communication between modules are encrypted transmission;Through AES-256 algorithm to the sensitive field stored in the database is symmetric encryption;Adopt the access control scheme based on the role, the user account is divided into administrator, quality data management engineer, operation and maintenance personnel different role;Key operation uses multi-factor authentication mechanism;The regular matching and keyword replacement processing of the confidential information of the enterprise in the user submitted instruction text and quality data.

10. The industrial quality data management system based on a conversational large language model of claim 9, wherein: Also includes performance optimization module, using containerized microservices deployment architecture, deployment of multiple parallel computing nodes;Through load balancing strategy will active session evenly distributed to different nodes;Adopt the asynchronous execution framework processing computationally intensive tasks;Through the WebSocket connection to the front end push progress update;Control system end to end average latency, wherein the instruction alignment module, semantic understanding reasoning, function module execution respectively in the preset time is completed.

Images