Fiber process intelligent generation method and system based on large language model

Abstract

The application provides a fiber process intelligent generation method and system based on a large language model, and belongs to the field of fiber material process design. The method comprises the following steps: designing standardized and analyzable fiber process coding rules; collecting fiber process original data, structurally splitting the original data according to the fiber process coding rules, and obtaining a fiber process training data set; training a preset fiber process large language model by using the fiber process training data set; and automatically generating fiber process codes meeting user requirements by using the trained fiber process large language model. The application breaks through the bottleneck of traditional fiber production relying on manual experience and lacking standardization, realizes the automatic generation of process schemes, and promotes the transformation of the fiber manufacturing industry to data-driven precision and intelligence.

Description

Technical Field

[0001] This invention belongs to the field of fiber material process design, and specifically relates to a method and system for intelligent generation of fiber processes based on a large language model, which is applicable to process design scenarios for various fiber materials such as carbon fiber and polymer synthetic fiber. Background Technology

[0002] The design of fiber material production processes has long faced the dual challenges of knowledge management and technological innovation. Traditional methods rely on expert experience to build process systems, which has three defects: (1) the risk of knowledge transmission gap, as the tacit knowledge of senior engineers is difficult to be effectively transformed into reusable process rules; (2) the degree of standardization of process documents is insufficient, and there are significant differences in the parameter systems and process descriptions adopted by different companies, resulting in a high failure rate of cross-platform process transfer; (3) the R&D cycle is lengthy, and the development of new processes requires a linear process of literature research, scheme design and trial production verification, which is time-consuming.

[0003] The existing method and system for quantitative prediction design of new materials assisted by large language model (CN118824432A) uses large language model to encode and map material composition and process parameters to generate performance prediction model to assist in the design of new materials. However, its core focuses on the construction of material performance prediction model and has not yet solved the problem of acquiring and preprocessing the whole chain of preparation data of high-quality fiber materials, including "composition-process-structure-performance / function". It lacks the ability to model the logical constraints of multiple processes in fiber production and it is difficult to realize the automatic generation of the whole process scheme.

[0004] In summary, large language models have demonstrated powerful cross-modal data processing capabilities in the Materials Genome Initiative, but their effective application in fiber process design has yet to materialize. Existing AI models primarily focus on molecular structure prediction, lacking the ability to model the logical constraints of the process chain. The generated process schemes often contain fundamental errors such as parameter conflicts and reversed steps, severely hindering the practical application of the technology. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a method and system for intelligent generation of fiber processes based on a large language model. This invention breaks through the bottlenecks of traditional fiber production, which relies on manual experience and lacks standardization, and realizes the automated generation of process solutions, promoting the transformation of the fiber manufacturing industry towards data-driven precision and intelligence.

[0006] A first aspect of this invention proposes a method for intelligent generation of fiber processes based on a large language model, comprising:

[0007] Design standardized and resolvable fiber process coding rules;

[0008] Collect raw fiber process data, and perform structured decomposition of the raw data according to the fiber process coding rules to obtain a fiber process training dataset;

[0009] A pre-defined fiber process large language model is trained using the aforementioned fiber process training dataset;

[0010] The trained fiber process language model is used to automatically generate fiber process codes that meet user requirements.

[0011] In one specific embodiment of the present invention, the fiber process coding rules include:

[0012] 1) The composition module describes the raw material formula required for fiber production;

[0013] 2) Process module, used to standardize the description of each processing step in fiber production, and to clarify the process type, parameter values, execution conditions and scope of each step;

[0014] 3) Structural module, used to record the microscopic and macroscopic structural features of the finished fiber product to verify that the structure of the process product meets expectations;

[0015] 4) Performance and Function Module: This module defines the various performance indicators and functional characteristics of fiber materials, specifies the testing standards, conditions, and effective range, and evaluates the application value of the processed products.

[0016] In one specific embodiment of the present invention, the fiber process coding rules also include anomaly handling principles;

[0017] The exception handling principles include:

[0018] When data is missing, it is marked as NA;

[0019] When contradictory data is found, retain the original data and mark it as ERR.

[0020] In one specific embodiment of the present invention, it further includes:

[0021] The raw data of the fiber process is processed using a large language model. During processing, preset prompts and the raw data are input into the large language model. The large language model decomposes the fiber preparation process contained in the raw data into component modules, process modules, structural modules, and performance and functional modules according to the fiber process coding rules, thereby generating training samples. All training samples constitute the fiber process training dataset.

[0022] In a specific embodiment of the present invention, the construction process of the fiber process large language model is as follows:

[0023] 1) The GPT-2 model is used as the base model;

[0024] The structure of GPT-2 is as follows:

[0025]

[0026]

[0027]

[0028] in, This is the initial hidden layer. This is the lth hidden layer in the middle. For the final number of attention heads, For the input word ID matrix, For word embedding matrix, For the position encoding matrix, For word probability distribution;

[0029] TransformerBlock represents the basic computational unit of the GPT-2 model. Each TransformerBlock consists of the following two parts in sequence: (1) Multi-head self-attention layer: This layer first processes the input vector The mapping is represented as a matrix of query Q, key K, and value V, and attention weights are calculated. ,in For Q and K vector dimensions, softmax transforms the similarity into a probability distribution; (2) Feedforward fully connected network: performs two-layer linear transformation and nonlinear activation on the vector at each position; finally, the output is processed by residual connection and layer normalization to obtain the hidden state of this layer. ;

[0030] 2) Expansion of the thesaurus in the field of fiber processing;

[0031] Candidate terms were selected from the fiber process training dataset. These candidate terms were specialized terms covering the entire fiber manufacturing process. A fiber domain saliency score was calculated for each candidate term.

[0032]

[0033] Where t represents any candidate word. The TF-IDF value of word t on the fiber technology training dataset is used to measure the importance of t in the fiber data. The TF-IDF value of word t in ordinary daily text is used to measure the frequency of t in daily corpora. IDF For inverse document word frequency, , ;

[0034] Judgment: If the fiber domain saliency score of any candidate term is greater than or equal to the set threshold, then the candidate term is determined to be a new fiber core term.

[0035] The N newly added fiber core terms obtained through the determination are merged into the original GPT-2 vocabulary to form a hybrid vocabulary;

[0036] 3) Initialize the word vectors of newly added terms;

[0037] The word vectors of the N newly added fiber domain terms selected in step 2) are initialized using a semantic nearest neighbor transfer and noise injection strategy. The specific method is as follows:

[0038] Calculate the semantic similarity between each newly added fiber domain term t and all words in the original vocabulary; based on semantic similarity, select the top n nearest neighbors with the highest similarity to t from the original vocabulary. ; Calculate the arithmetic mean of the nearest neighbor word vectors: ,in The original GPT-2 embedding matrix is ​​used; Gaussian noise is superimposed to generate word vectors for the newly added fiber domain term t:

[0039]

[0040] in, Superimposed Gaussian noise ;

[0041] The word vectors corresponding to the N newly added fiber core terms are merged with the original GPT-2 embedding matrix to form a new embedding matrix;

[0042] 4) Process-aware location coding;

[0043] Inject process timing features into the model's location encoding layer and encode process stage identifiers as additional location bias terms;

[0044]

[0045]

[0046] Where pos is the absolute position number of the word in the sequence, i is the dimension sharding index, and d is the hidden layer dimension; These are the stage feature weight coefficients. For learnable projection matrices, This is a process stage identifier vector, using static encoding. .

[0047] In a specific embodiment of the present invention, the training of the preset fiber technology large language model using the fiber technology training dataset adopts a hierarchical progressive fine-tuning training method, specifically including:

[0048] Phase 1: Fine-tuning of newly added word vectors;

[0049] In this phase, the base model is frozen, and the original parameters of GPT-2 remain unchanged. The semantic embedding of the newly added fiber domain terms is ensured to be stable by training the word vector parameters of the newly added fiber domain terms.

[0050] Phase Two: Unfreezing of High-Level Transformers;

[0051] In this stage, the last four Transformer layers of the GPT-2 model are unfrozen, and word vectors of newly added fiber domain terms are trained to adjust the model's ability to adjust high-level semantic composition.

[0052] Phase 3: Fine-tuning of the entire model;

[0053] This stage involves unfreezing all model parameters and performing global optimization.

[0054] In one specific embodiment of the present invention, it further includes:

[0055] The fiber process codes output by the trained fiber process big language model are dynamically verified.

[0056] The dynamic verification process includes checking for spelling errors, parameter errors, and format inconsistencies in the process flow in the fiber process code, and then converting the fiber process code into a natural language description.

[0057] A second aspect of this invention provides a fiber process intelligent generation device based on a large language model, comprising:

[0058] The fiber process coding rule design module is used to design standardized and parsable fiber process coding rules.

[0059] The training dataset construction module is used to collect raw fiber process data and perform structured splitting of the raw data according to the fiber process coding rules to obtain a fiber process training dataset.

[0060] The fiber process large language model training module is used to train a preset fiber process large language model using the fiber process training dataset.

[0061] The fiber process coding generation module is used to automatically generate fiber process codes that meet user requirements using the trained fiber process large language model.

[0062] A third aspect of the present invention provides an electronic device comprising:

[0063] At least one processor; and a memory communicatively connected to said at least one processor;

[0064] The memory stores instructions that can be executed by the at least one processor, and the instructions are configured to execute the above-described intelligent generation method for fiber processes based on a large language model.

[0065] To implement the above embodiments, a fourth aspect of the present invention provides a computer-readable storage medium storing computer instructions for causing the computer to execute a fiber process intelligent generation method based on a large language model.

[0066] Features and beneficial effects of the present invention:

[0067] 1. Break through the bottlenecks of traditional process design and solve the problems of knowledge gaps and standardization.

[0068] This invention's pioneering Fiber Process Coding (FPC) rules, defined through a four-module structure of "composition-process-structure-performance," transform tacit expert knowledge into a standardized parameter system that can be parsed by machines. This unifies the cross-enterprise process description format and solves the problems of knowledge transfer reliance on manual intervention and chaotic parameter systems leading to transplantation failures in traditional methods. Through dynamic range parameters and nested sub-process design, it is compatible with complex process scenarios, significantly reducing the cost of technology reuse for enterprises.

[0069] 2. Construct a domain-enhanced large model to achieve precise process chain logic control.

[0070] By expanding the domain vocabulary, employing process-aware positional encoding, and using hierarchical progressive fine-tuning techniques, this invention enables a general-purpose large language model to possess the ability to model fiber process chain logic. The model can autonomously identify raw material fluctuations and dynamically adjust process parameters, while also supporting the accurate generation of nested processes, improving the professionalism of the generated content and the stability of the process. This solves the practicality obstacles of existing AI models, such as process inversion and parameter incompatibility.

[0071] 3. Drive the intelligent upgrading of fiber manufacturing and empower efficient collaboration across the entire industry chain.

[0072] The successful application of this invention will significantly shorten the R&D cycle for fiber process optimization and spinning parameter design, and reduce trial production and verification costs. Through a two-way conversion mechanism between FPC coding and natural language, it can automatically output process documents that conform to industry standards, reducing manual workload and providing standardized process interfaces for smart factories, enabling multi-device parameter linkage optimization. It is expected to generate a market application worth hundreds of millions in high-end materials fields such as aerospace and biomedical fibers, driving the industry's transformation from "experience-based trial and error" to "data-driven" approaches. Attached Figure Description

[0073] Figure 1 This is a flowchart of a fiber process intelligent generation method based on a large language model, according to an embodiment of the present invention. Detailed Implementation

[0074] This invention proposes a method and system for intelligent generation of fiber processes based on a large language model. The invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0075] A first aspect of this invention proposes a method for intelligent generation of fiber processes based on a large language model, comprising:

[0076] Design standardized and resolvable fiber process coding rules;

[0077] Collect raw fiber process data, and perform structured decomposition of the raw data according to the fiber process coding rules to obtain a fiber process training dataset;

[0078] A pre-defined fiber process large language model is trained using the aforementioned fiber process training dataset;

[0079] The trained fiber process language model is used to automatically generate fiber process codes that meet user requirements.

[0080] In a specific embodiment of the present invention, the overall process of the intelligent generation method for fiber processes based on a large language model is as follows: Figure 1 As shown, it includes:

[0081] 1) Design fiber process coding rules.

[0082] In this embodiment, based on the entire chain of fiber material preparation data (composition-process-structure-performance / function), a standardized, machine-resolvable coding rule, Fiber Process Code (FPC), is designed to describe the fiber production process. The FPC coding rule is defined as follows:

[0083] a) Composition module (Comp).

[0084] The composition module is used to comprehensively describe the raw material formula required for fiber production, including: main materials and their content, auxiliary materials and their addition order, material state, concentration range, etc. It supports nested sub-processes to define complex raw material synthesis paths, ensuring that material information is complete, clear and parsable.

[0085] Rule description:

[0086] Main material definition: identified by [Comp], in the format "main material=content:unit", for example "PAN=88%".

[0087] Nested sub-processes: Identifiers used in the production process of the main material ... A package or subprocess may contain other modules (such as process steps).

[0088] Auxiliary materials: The auxiliary materials are separated from the main materials by a semicolon ";", and the content of the auxiliary materials is described.

[0089] Special handling:

[0090] Range value: Use "-" to indicate the concentration range, such as "Ag3PO4=10-30%".

[0091] Material flow direction: The order of addition is indicated by the arrow symbol "→", for example, "AgNO3→H2O" means silver nitrate is added to water.

[0092] Additional attributes: Specify the state or condition in parentheses after the material, such as "PLA (particulate state) = 2g".

[0093] Example:

[0094] [Comp]

[0095] PAN=88%

[0096]

[0097] [Comp]AN=95%; AIBN=5%

[0098] <step> [Proc-Type=聚合]Temp=60:℃(N2); Time=24:h < / step>

[0099]

[0100] DMSO=10%; CNT=2

[0101] Meaning:

[0102] Initiator AIBN (5%) was synthesized via polymerization (60°C, nitrogen protection, 24 hours). Auxiliary materials: solvent DMSO (10%) and carbon nanotubes (CNT, 2wt%) were added.

[0103] b) Process module (Proc).

[0104] The process module is used to standardize the description of each processing step in fiber production, clarify the process type, parameter values, execution conditions and range of each step, support the logical sequence modeling of complex multi-step processes, and ensure that the process chain is reasonable and conflict-free.

[0105] Rule description:

[0106] Process steps: using identifiers <step> ...< / step> Each step in the package begins with [Proc-Type=Process Type], for example, [Proc-Type=Electrospinning].

[0107] Parameter definition: The format is parameter name=value:unit (condition), for example, "Voltage=20:kV(generator, ±5%)".

[0108] Special handling:

[0109] Range parameter: Use a hyphen "-" to indicate the parameter range, such as "Temp=180-220:℃".

[0110] Multiple conditions can be added within parentheses, such as "FlowRate=1.2:ml / h (injection pump, accuracy 0.1%)".

[0111] Example:

[0112] <step>

[0113] [Proc-Type=Electrospinning]

[0114] Voltage = 20 kV (generator, ±5%)

[0115] FlowRate = 1.2 ml / h (infusion pump)

[0116] Collector = Aluminum foil (rotating, 500 rpm)

[0117] < / step>

[0118] <step>

[0119] [Proc-Type=Post-Stretch]

[0120] Ratio=3.5X; Speed=10 mm / min

[0121] < / step>

[0122] Meaning:

[0123] Electrospinning steps: voltage 20kV (±5% fluctuation allowed), flow rate 1.2ml / h (controlled by syringe pump), aluminum foil collector rotating at 500rpm. Post-stretching steps: stretching ratio 3.5 times, stretching speed 10mm / min.

[0124] c) Struct module.

[0125] The structural module is used to record the microscopic and macroscopic structural characteristics of the finished fiber product, including indicators such as size, porosity, orientation, and crystallinity, as well as their measurement methods and statistical properties, to verify that the structure of the processed product meets expectations.

[0126] Rule description:

[0127] Parameter definition: Identified by [Struct], in the format "parameter name=value:unit(method=measurement method, additional information)", for example "FiberDia=500:nm(method=SEM, SD=50)".

[0128] Special handling:

[0129] Distribution description: If the aperture distribution is Gaussian, it is labeled "Pore=30-100:nm(distribution=Gaussian)".

[0130] Example:

[0131] [Struct]

[0132] FiberDia=500:nm(method=SEM, SD=50)

[0133] Pore=30-100:nm(distribution=Gaussian)

[0134] Orientation=75° (method=XRD)

[0135] Crystallinity = 45% (method = DSC)

[0136] Meaning:

[0137] Fiber diameter: 500 nm (measured by scanning electron microscopy, standard deviation 50 nm). Pore size: 30-100 nm (Gaussian distribution). Orientation: 75° (measured by X-ray diffraction). Crystallinity: 45% (measured by differential scanning calorimetry).

[0138] d) Performance and functional modules (Perf / Func).

[0139] The performance and function module is used to define the mechanical, electrical, antibacterial, and UV-resistant properties and functional characteristics of fiber materials, specify the test standards, conditions, and effective range, and is used to evaluate the application value of the process products.

[0140] Rule description:

[0141] Performance parameters: identified by [Perf], in the format "parameter name=value:unit(test standard)", for example "Tensile=120:MPa(standard=ASTM-D638)".

[0142] Functional attributes: identified by [Func], in the format "attribute name=value(condition)", for example "UV-Resist=√(range=300-400:nm)".

[0143] Example:

[0144] [Perf]

[0145] Tensile=120:MPa(standard=ASTM-D638)

[0146] Conductivity=1e3:S / m(25:℃, RH=45%)

[0147] [Func]

[0148] Antibacterial = 99.9% (strain = E. coli)

[0149] UV-Resist=√(range=300-400:nm)

[0150] Meaning:

[0151] Mechanical properties: Tensile strength 120 MPa (ASTM standard test). Electrical conductivity: 1 × 10³ S / m (25℃, 45% relative humidity). Antibacterial properties: 99.9% inhibition rate against Escherichia coli. UV resistance: Effective in the 300-400 nm wavelength range.

[0152] Furthermore, the FPC encoding rules also include:

[0153] Exception handling principles:

[0154] Missing data: When missing data occurs, it is marked as NA, for example, [Perf] Conductivity=NA(test=not tested).

[0155] Conflicting data: When conflicting data occurs, preserve the original data and mark it as ERR, for example, [Struct] FiberDia=200:nm / 250:nm(ERR, method=SEM / AFM).

[0156] The meanings of the symbols can be described in the following table:

[0157] Table 1. Meaning of symbols in FPC rules in the method described in this embodiment.

[0158] symbol name Application scenarios Example ; Parameter separator Separate different elements within the same module / step Voltage=20:kV;FlowRate=1.2:ml / h → Material flow symbol Indicates the order of addition (A→B means A is added to B). AgNO3→H2O () Condition specifier Add supplementary explanations or limiting conditions for parameters. Temp=60:℃(N2)Conductivity=1e3:S / m(25:℃) : Unit separator Separate numerical values ​​from units Time=3:hFiberDia=500:nm - Scope marker Indicates the parameter range or gradient Temp = 180-220℃; Ag3PO4 = 10-30% = Assignment operator Connect parameter name and value

[0159] In this embodiment, the FPC encoding rules transform complex fiber production processes into a machine-readable format through modular and standardized parameter definitions. They support nested processes, dynamic range parameters, and anomaly annotations, providing a structured data foundation for subsequent AI model training and process generation.

[0160] 2) Based on the coding rules established in step 1), data collection and structured processing are performed to obtain the fiber process training dataset.

[0161] In this embodiment, data is collected from fiber technology-related literature, mainly including patent literature, academic papers, and process manuals in the fiber field. The formats include pdf, txt, doc, docx, etc. Specifically, the method is to download fiber technology-related PDF / Word documents from patent databases (such as CNKI, Espacenet) and academic platforms (such as CNKI, ScienceDirect), or to organize existing experimental record documents.

[0162] The collected data is structured using common general-purpose large language model tools. This can be achieved by calling online APIs such as OpenAI GPT and Deepseek, specifically by registering an official platform account, obtaining API access permissions and keys, installing the API call library, and writing Python processing scripts. Alternatively, the official website of the large language model can be used for dialogue processing. Specifically, specific prompt words and documents are input simultaneously into the general-purpose large language model, which outputs data in the appropriate format according to the prompt words. A structured extraction instruction template, Doc2FPC-prompt, needs to be designed to support the extraction of fiber process-related structural data from the literature data and convert it into a standard format suitable for subsequent model training. During this process, the fiber preparation process in the original text is decomposed into four modules according to FPC encoding rules: composition, process, structure, and performance and function. In this embodiment, the collected original literature data must contain a complete process chain description, such as material formulation, processing steps, product structure and performance indicators, and the terminology must be standardized. The core characteristics of the processed data include standardized parameter formats, modular classification, and nested hierarchical structures.

[0163] In one specific embodiment of the present invention, the structured extraction instruction Doc2FPC-prompt is designed as follows:

[0164] You are a polymer materials expert and need to extract fiber process codes (FPCs) from documents, strictly adhering to the following rules:

[0165] Synonym consistency: Terms with the same meaning should be expressed consistently.

[0166] Single Document Single FPC: Only one complete FPC code is extracted from each document.

[0167] Output FPC only: Outputs standardized FPC text directly without adding any explanatory content.

[0168] The Fiber Process Code (FPC) is as follows: (The fiber process code from step 1 is added here.)

[0169] Through the above steps, this embodiment will obtain a large amount of FPC structured data based on the number of original documents. The data will be in JSON format, categorized into four modules, covering fields such as material ratio, process parameters, fiber morphology, and mechanical properties. The training set in this embodiment must cover at least 1,000 training samples, encompassing mainstream process types such as carbon fiber and nanofiber. For special scenarios (such as biomedical fibers), additional samples are needed to bring the total to over 2,000 to ensure the model's generalization ability to long-tailed data. Simultaneously, the sample ratio for different process types must be balanced; for example, avoid having too much spinning process data and too little heat-setting data to prevent model bias.

[0170] The training samples in a specific embodiment of the present invention are as follows:

[0171] [Comp]PAN=88% [Comp]AN=95%; AIBN=5% <step> [Proc-Type=聚合]Temp=60:℃(N2);Time=24:h < / step> DMSO=10%; CNT=2wt%

[0172] <step> [Proc-Type=Electrospinning]Voltage=20:kV(Generator, ±5%); FlowRate=1.2:ml / h(SyringePump); Collector=Al-foil(Rotating, 500:rpm)< / step>

[0173] <step> [Proc-Type=Post-Stretch]Ratio=3.5X; Speed=10:mm / min< / step>

[0174] [Struct]FiberDia=500:nm(method=SEM,SD=50);Pore=30-100:nm(distribution=Gaussian);Orientation=75°(method=XRD);Crystallinity=45%(method=DSC)

[0175] [Perf]Tensile=120:MPa(standard=ASTM-D638);Conductivity=1e3:S / m(25:℃,RH=45%)

[0176] [Func]Antibacterial=99.9%(strain=E.coli);UV-Resist=√(range=300-400:nm)

[0177] Explanation: This fiber process code (FPC) describes the preparation process of the PAN-based material: the main material is PAN (88%), and its synthesis sub-process is achieved through a polymerization reaction (95% PAN monomer, 5% initiator AIBN, reaction at 60℃ under nitrogen protection for 24 hours), with the addition of solvent DMSO (10%) and carbon nanotubes (CNT, 2wt%). The process module includes two stages: electrospinning and post-stretching. Electrospinning uses a voltage of 20kV (±5% fluctuation, generator controlled), a flow rate of 1.2ml / h (driven by a syringe pump), and an aluminum foil collector rotating at 500rpm. The post-stretching process is set with a stretch ratio of 3.5 and a stretching speed of 10mm / min. The structural characteristics of the resulting fiber include a diameter of 500nm (SEM measurement, standard deviation 50nm), a pore size of 30-100nm (Gaussian distribution), a molecular chain orientation angle of 75° (XRD determination), and 45% crystallinity (DSC analysis). Performance tests show that the tensile strength reaches 120MPa (ASTM-D638 standard), the conductivity is 1e3 S / m (25℃ / 45% humidity environment), and it also has a 99.9% E. coli inhibition rate and 300-400nm ultraviolet protection function.

[0178] 3) Construct a large language model for fiber processing.

[0179] The fiber processing language model in this embodiment is based on a generative pre-trained transformer (GPT) architecture, consisting of the following parts: 1. Embedding layer: Input words are mapped to word vectors through an embedding matrix; 2. Position encoding layer: On top of standard absolute position encoding, a process-aware position bias term is superimposed to enhance the model's understanding of the process stage sequence; 3. Transformer encoding layer: A multi-layer transformer of GPT itself, including multi-head self-attention, residual connections, feedforward neural networks, etc., used to model sequence context and generation logic; 4. Output layer: The transformer output is mapped back to the vocabulary dimension to generate a probability distribution. The system streams the next word based on the probability distribution of the current time step. The specific steps of model construction are as follows:

[0180] 3-1) Select the base model of the fiber process large language model.

[0181] In this embodiment, the GPT-2 series models are used as the base models for the fiber process large language model. The specific model size is selected based on the actual amount of JSON data acquired, according to the following rules: if the number of training set samples is between 1k and 5k, the GPT-2 Small model with a hidden layer dimension of 768 and 12 attention heads is selected as the base model; if the number of training set samples is between 5k and 20k, the GPT-2 Medium model with a hidden layer dimension of 1024 and 16 attention heads is selected as the base model; if the number of training set samples is between 20k and 100k, the GPT-2 Large model with a hidden layer dimension of 1280 and 20 attention heads is selected as the base model; and if the number of training set samples is more than 100k, the GPT-2 XL model with a hidden layer dimension of 1600 and 25 attention heads is selected as the base model.

[0182] The structure of GPT-2 is as follows:

[0183]

[0184]

[0185]

[0186] in, This is the initial hidden layer. This is the lth hidden layer in the middle. For the final number of attention heads, For the input word ID matrix, For word embedding matrix, For the position encoding matrix, The probability distribution is denoted by . Here, TransformerBlock represents the basic computational unit of the GPT-2 model. Each TransformerBlock consists of the following two parts in sequence: (1) Multi-head self-attention layer. This layer first processes the input vector... The mapping is represented as a matrix of query Q, key K, and value V, and attention weights are calculated. ,in For Q and K vector dimensions, softmax transforms the similarity into a probability distribution. (2) Feedforward fully connected network. Two layers of linear transformation and nonlinear activation are performed on the vector at each position. Finally, the output is processed through residual connections and layer normalization to obtain the hidden state of this layer. .

[0187] 3-2) Expansion of the thesaurus in the field of fiber processing.

[0188] This embodiment uses statistical analysis and domain knowledge to screen high-frequency technical terms in fiber processing. The specific method is as follows:

[0189] First, terminology was screened for the structural data related to fiber manufacturing processes. The candidate terms covered professional terms for the entire fiber manufacturing process, mainly including five categories: materials, processes, parameters, structure, and properties. The materials category involved main materials (such as PAN and PLA), solvents (such as DMSO and DMF), and functional components (such as AgNPs and GO). The processes category included process types such as electrospinning and pre-oxidation, as well as equipment terms such as spinning heads and high-voltage generators. The parameters category covered process parameters such as voltage and flow rate, and physicochemical indicators such as viscosity and molecular weight. The structure and properties category focused on characteristics such as fiber diameter and tensile strength, as well as testing methods such as SEM and XRD. In addition, special symbols and units such as N2 protection and rpm were also included.

[0190] In this embodiment, the candidate words are derived from the fiber process training dataset obtained in step 2). Candidate words can be selected manually or all segmented words can be used directly after word segmentation. The fiber domain saliency score of the candidate words is calculated.

[0191]

[0192] Where t represents any candidate word. The TF-IDF value of word t on the fiber technology training dataset is used to measure the importance of t in the fiber data. The TF-IDF value of word t in ordinary everyday text (such as a general corpus) is used to measure the frequency of t in everyday text. IDF For inverse document word frequency, , The more frequently a word appears in the current training samples (higher TF) and the less frequently it appears in all training samples (higher IDF), the larger its TF-IDF, and the more important the word is. Therefore, if a word is important for fiber processing data... The larger the document, the better for ordinary daily documents. The smaller the value, the greater its significance in the fiber domain.

[0193] In one specific embodiment of the present invention, a threshold of Score ≥ 0.8 is set. If a word scores high in some samples and low in some samples, the system uses the mean of Score(t) calculated from the entire training set to judge and exclude common words. Finally, the fiber field experts manually review and retain the number of newly added fiber-related terms as N. The original GPT-2 vocabulary (50257 words) and the newly added terms are merged using SentencePiece to form a mixed vocabulary of (50257+N) words.

[0194] 3-3) Initialize the word vectors of new terms.

[0195] In this embodiment, the word vectors of the N newly added fiber domain terms selected in step 3-2) are initialized using a semantic nearest neighbor migration and noise injection strategy. The specific method is as follows:

[0196] The semantic similarity between each newly added fiber domain term t and all words in the original GPT-2 vocabulary (50257 words in total) is calculated using pre-trained models such as Sentence-BERT. Based on the semantic similarity of all words, the top n nearest neighbors with the highest similarity to t are selected from the original vocabulary. In this embodiment, the cosine similarity is required to exceed a set similarity threshold, which is 0.7 in this embodiment; and n=5 in this embodiment. The arithmetic mean of the nearest neighbor word vectors is calculated. ,in This is the original GPT-2 embedding matrix. Gaussian noise is then added to generate word vectors for the newly added fiber domain term t:

[0197]

[0198] Superimposed Gaussian noise It can provide fine-tuning degrees of freedom while preserving semantic relevance. The word vectors corresponding to the N newly added fiber core terms are merged with the original GPT-2 embedding matrix to form a new embedding matrix of (50257+N)×768 dimensions.

[0199] 3-4) Process sensing location coding.

[0200] In fiber processing generation tasks, the stage characteristics of processes (such as raw material processing → spinning → post-processing) are more critical than absolute position. Traditional positional encoding only captures word order and cannot distinguish the positional relationships within different process stages. This embodiment injects process temporal features into the positional encoding layer and encodes process stage identifiers as additional positional bias terms, enhancing the model's ability to perceive the process chain sequence.

[0201]

[0202]

[0203] Where pos is the absolute position index of the word in the sequence, i is the dimension shard index, and d is the hidden layer dimension, which is 768 for GPT-2. These are the stage feature weight coefficients. For learnable projection matrices, This is a process stage identifier vector, using static encoding. The value of k can be defined by the user based on the complexity of the process.

[0204] 4) Use the fiber process training dataset obtained in step 2) to perform hierarchical progressive fine-tuning on the fiber process large language model established in step 3) to obtain the trained fiber process large language model.

[0205] In this embodiment, the initialization model based on the domain vocabulary and process-aware coding is subjected to a three-stage lightweight fine-tuning: Stage 1 freezes the base model, keeping the original GPT-2 parameters (including the original word vectors, attention layer, feedforward neural network (FFN) etc.) unchanged, and only training the newly added word vectors; Stage 2 unfreezes the high-level transformer, adjusting the high-level semantic combination capability to adapt to the complex logic of fiber process; Stage 3 fine-tunes the entire model, globally optimizing the parameter distribution and balancing domain adaptation and general language capability.

[0206] Specifically, Phase 1: Fine-tuning of newly added word vectors;

[0207] This phase only unfreezes the word vector parameters of newly added fiber domain terms related to the fiber domain vocabulary, keeping the original GPT-2 parameters (such as attention layers, FFN, etc.) frozen to ensure the generality of the base model is not compromised. Training uses a low learning rate (lr=5e-5) and the AdamW optimizer (betas=(0.9, 0.98), weight decay 0.01), with a batch size of 32 to accommodate memory limitations. The training cycle is fixed at 10-20 rounds, or terminated early according to the early stopping mechanism (no decrease in validation loss for 3 consecutive rounds). The core termination conditions include a validation set domain term recall rate (such as keywords like "electrospinning" and "PAN") of over 95%, and the cosine similarity fluctuation of the newly added word vectors converging to a difference of less than 0.01 between adjacent rounds to ensure the semantic embedding of the new words is stable.

[0208] Phase Two: Unfreezing of High-Level Transformers (Top-4 Layers);

[0209] This phase involves unfreezing the last four Transformer layers of the GPT-2 model (e.g., unfreezing layers 8-11 when the total number of layers is 12) and adding new word vectors. The aim is to adjust the high-level semantic composition capabilities to adapt to the complex logic of fiber processing. A hierarchical learning rate strategy is used for training parameters: the learning rate for high-level Transformer layers is set to 1e-4, while the learning rate for word vector layers is reduced to 3e-5. The optimizer remains AdamW (betas=(0.9, 0.999), with a weight decay of 0.1). The batch size is dynamically increased to 64 to utilize the unfrozen parameter capacity, and the training cycle is extended to 20-50 epochs. The early stopping condition is set to ensure that the validation loss does not significantly decrease for five consecutive epochs. The termination criterion requires meeting the logical coherence of the process chain (e.g., the correct generation accuracy of the "polymerization → spinning → post-stretching" process exceeds 90%).

[0210] Phase 3: Fine-tuning of the entire model;

[0211] In the final stage, all model parameters (including embeddings, attention layers, FFN, etc.) are unfrozen, and global optimization is performed to balance domain adaptability and general capabilities. The learning rate is further reduced to 3e-6 to avoid catastrophic forgetting, and the LAMB optimizer is used to support large-batch training (batch size=128), with strong regularization introduced (Dropout rate 0.2, gradient pruning threshold 1.0). The training cycle is extended to 50-100 epochs, and the early stopping mechanism is set to verify that the loss has not improved for 10 consecutive epochs. The termination condition requires that the domain task metrics (such as the accuracy of process parameter generation, F1 score) reach a preset threshold (e.g., F1 ≥ 0.92). During training, cosine annealing learning rate scheduling is used (cycle = 100, minimum learning rate 1e-7), and the performance of domain tasks and general tasks is mixed and verified in each epoch to ensure that the model is accurate in fiber process scenarios while retaining basic language understanding capabilities.

[0212] 5) Intelligent generation of fiber processes.

[0213] In this step, based on the Fiber Processing Language Model (FiberGPT) trained in step 4) and the standard process framework, the system will automatically generate fiber production processes that meet the user's needs. The user inputs specific performance or functional requirements, process conditions, and the characteristics of the target material. FiberGPT will then automatically generate the corresponding fiber process code (FPC) based on these inputs. These process flows include multiple aspects such as raw material selection, process steps, structure, and performance, ensuring that the generated process plan meets actual production requirements.

[0214] In this embodiment, the user input can specify performance goals, process constraints, and material characteristics. The system will then parse and generate a complete FPC code that includes component selection, spinning parameters, and post-processing conditions. For example: "Please generate a production process for antibacterial nanofibers suitable for medical dressings, requiring the fiber diameter to be controlled within the range of 200-300nm."

[0215] Optionally, the preliminary process code (FPC) generated in this embodiment can be further optimized through dynamic verification. The dynamic verification process includes checking for spelling errors, parameter errors, and format inconsistencies in the process flow, ensuring the accuracy and clarity of the process description. It is then converted into natural language descriptions that conform to industry standards, facilitating engineer understanding and execution. Specifically, online APIs such as GPT-o1 and Deepseek-R1, along with FPC2Doc-prompt prompts, are used for translation and natural language optimization to automatically generate standard-compliant fiber production process documents. These documents not only describe each process step in detail but also provide operational requirements, experimental conditions, and other detailed information, ensuring strict adherence to every step of the production process. The process flow is ultimately output in a standardized report format, which can be directly applied to production or used for subsequent process optimization and adjustments.

[0216] In one specific embodiment of the present invention, the FPC2Doc-prompt prompt is as follows:

[0217] You are an expert in polymer materials. You need to convert Fiber Processing Code (FPC) into a fiber production process described in natural Chinese, according to the following rules. Only output the complete fiber production process; do not include any other content.

[0218] To implement the above embodiments, a second aspect of the present invention proposes a fiber process intelligent generation device based on a large language model, comprising:

[0219] The fiber process coding rule design module is used to design standardized and parsable fiber process coding rules.

[0220] The training dataset construction module is used to collect raw fiber process data and perform structured splitting of the raw data according to the fiber process coding rules to obtain a fiber process training dataset.

[0221] The fiber process large language model training module is used to train a preset fiber process large language model using the fiber process training dataset.

[0222] The fiber process coding generation module is used to automatically generate fiber process codes that meet user requirements using the trained fiber process large language model.

[0223] It should be noted that the foregoing explanation of an embodiment of a fiber process intelligent generation method based on a large language model also applies to a fiber process intelligent generation device based on a large language model in this embodiment, and will not be repeated here. According to an embodiment of the present invention, a fiber process intelligent generation device based on a large language model designs standardized and parsable fiber process coding rules; collects raw fiber process data; performs structured decomposition of the raw data according to the fiber process coding rules to obtain a fiber process training dataset; trains a preset fiber process large language model using the fiber process training dataset; and automatically generates fiber process codes that meet user requirements using the trained fiber process large language model. This overcomes the bottleneck of traditional fiber production relying on manual experience and lacking standardization, achieving automated generation of process solutions and promoting the fiber manufacturing industry's transformation towards data-driven precision and intelligence.

[0224] In one specific embodiment of the present invention, the fiber process coding rules include:

[0225] 1) The composition module describes the raw material formula required for fiber production;

[0226] 2) Process module, used to standardize the description of each processing step in fiber production, and to clarify the process type, parameter values, execution conditions and scope of each step;

[0227] 3) Structural module, used to record the microscopic and macroscopic structural features of the finished fiber product to verify that the structure of the process product meets expectations;

[0228] 4) Performance and Function Module: This module defines the various performance indicators and functional characteristics of fiber materials, specifies the testing standards, conditions, and effective range, and evaluates the application value of the processed products.

[0229] In one specific embodiment of the present invention, the fiber process coding rules also include anomaly handling principles;

[0230] The exception handling principles include:

[0231] When data is missing, it is marked as NA;

[0232] When contradictory data is found, retain the original data and mark it as ERR.

[0233] In one specific embodiment of the present invention, it further includes:

[0234] The raw data of the fiber process is processed using a large language model. During processing, preset prompts and the raw data are input into the large language model. The large language model decomposes the fiber preparation process contained in the raw data into component modules, process modules, structural modules, and performance and functional modules according to the fiber process coding rules, thereby generating training samples. All training samples constitute the fiber process training dataset.

[0235] In a specific embodiment of the present invention, the construction process of the fiber process large language model is as follows:

[0236] 1) The GPT-2 model is used as the base model;

[0237] The structure of GPT-2 is as follows:

[0238]

[0239]

[0240]

[0241] in, This is the initial hidden layer. This is the lth hidden layer in the middle. For the final number of attention heads, For the input word ID matrix, For word embedding matrix, For the position encoding matrix, For word probability distribution;

[0242] TransformerBlock represents the basic computational unit of the GPT-2 model. Each TransformerBlock consists of the following two parts in sequence: (1) Multi-head self-attention layer: This layer first processes the input vector The mapping is represented as a matrix of query Q, key K, and value V, and attention weights are calculated. ,in For Q and K vector dimensions, softmax transforms the similarity into a probability distribution; (2) Feedforward fully connected network: performs two-layer linear transformation and nonlinear activation on the vector at each position; finally, the output is processed by residual connection and layer normalization to obtain the hidden state of this layer. ;

[0243] 2) Expansion of the thesaurus in the field of fiber processing;

[0244] Candidate terms were selected from the fiber process training dataset. These candidate terms were specialized terms covering the entire fiber manufacturing process. A fiber domain saliency score was calculated for each candidate term.

[0245]

[0246] Where t represents any candidate word. The TF-IDF value of word t on the fiber technology training dataset is used to measure the importance of t in the fiber data. The TF-IDF value of word t in ordinary daily text is used to measure the frequency of t in daily corpora. IDF For inverse document word frequency, , ;

[0247] Judgment: If the fiber domain saliency score of any candidate term is greater than or equal to the set threshold, then the candidate term is determined to be a new fiber core term.

[0248] The N newly added fiber core terms obtained through the determination are merged into the original GPT-2 vocabulary to form a hybrid vocabulary;

[0249] 3) Initialize the word vectors of newly added terms;

[0250] The word vectors of the N newly added fiber domain terms selected in step 2) are initialized using a semantic nearest neighbor transfer and noise injection strategy. The specific method is as follows:

[0251] Calculate the semantic similarity between each newly added fiber domain term t and all words in the original vocabulary; based on semantic similarity, select the top n nearest neighbors with the highest similarity to t from the original vocabulary. ; Calculate the arithmetic mean of the nearest neighbor word vectors: ,in The original GPT-2 embedding matrix is ​​used; Gaussian noise is superimposed to generate word vectors for the newly added fiber domain term t:

[0252]

[0253] in, Superimposed Gaussian noise ;

[0254] The word vectors corresponding to the N newly added fiber core terms are merged with the original GPT-2 embedding matrix to form a new embedding matrix;

[0255] 4) Process-aware location coding;

[0256] Inject process timing features into the model's location encoding layer and encode process stage identifiers as additional location bias terms;

[0257]

[0258]

[0259] Where pos is the absolute position number of the word in the sequence, i is the dimension sharding index, and d is the hidden layer dimension; These are the stage feature weight coefficients. For learnable projection matrices, This is a process stage identifier vector, using static encoding. .

[0260] In a specific embodiment of the present invention, the training of the preset fiber technology large language model using the fiber technology training dataset adopts a hierarchical progressive fine-tuning training method, specifically including:

[0261] Phase 1: Fine-tuning of newly added word vectors;

[0262] In this phase, the base model is frozen, and the original parameters of GPT-2 remain unchanged. The semantic embedding of the newly added fiber domain terms is ensured to be stable by training the word vector parameters of the newly added fiber domain terms.

[0263] Phase Two: Unfreezing of High-Level Transformers;

[0264] In this stage, the last four Transformer layers of the GPT-2 model are unfrozen, and word vectors of newly added fiber domain terms are trained to adjust the model's ability to adjust high-level semantic composition.

[0265] Phase 3: Fine-tuning of the entire model;

[0266] This stage involves unfreezing all model parameters and performing global optimization.

[0267] In one specific embodiment of the present invention, it further includes:

[0268] The fiber process codes output by the trained fiber process big language model are dynamically verified.

[0269] The dynamic verification process includes checking for spelling errors, parameter errors, and format inconsistencies in the process flow in the fiber process code, and then converting the fiber process code into a natural language description.

[0270] To implement the above embodiments, a third aspect of the present invention provides an electronic device, comprising:

[0271] At least one processor; and a memory communicatively connected to said at least one processor;

[0272] The memory stores instructions that can be executed by the at least one processor, and the instructions are configured to execute the above-described intelligent generation method for fiber processes based on a large language model.

[0273] To implement the above embodiments, a fourth aspect of the present invention provides a computer-readable storage medium storing computer instructions for causing the computer to execute a fiber process intelligent generation method based on a large language model.

[0274] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0275] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform a fiber process intelligent generation method based on a large language model according to the above embodiments.

[0276] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0277] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0278] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0279] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0280] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0281] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0282] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0283] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0284] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for intelligent generation of fiber processes based on a large language model, characterized in that, include: Design standardized and resolvable fiber process coding rules; Collect raw fiber process data, and perform structured decomposition of the raw data according to the fiber process coding rules to obtain a fiber process training dataset; A pre-defined fiber process large language model is trained using the aforementioned fiber process training dataset; The trained fiber process language model is used to automatically generate fiber process codes that meet user requirements. The fiber process coding rules include: 1) The composition module describes the raw material formula required for fiber production; 2) Process module, used to standardize the description of each processing step in fiber production, and to clarify the process type, parameter values, execution conditions and scope of each step; 3) Structural module, used to record the microscopic and macroscopic structural features of the finished fiber product to verify that the structure of the process product meets expectations; 4) Performance and Function Module: This module defines the various performance indicators and functional characteristics of fiber materials, specifies the test standards, conditions, and effective range, and evaluates the application value of the processed products. The construction process of the fiber process large language model is as follows: 1) The GPT-2 model is used as the base model; The structure of GPT-2 is as follows: in, This is the initial hidden layer. This is the lth hidden layer in the middle. For the final number of attention heads, For the input word ID matrix, For word embedding matrix, For the position encoding matrix, For word probability distribution; TransformerBlock represents the basic computational unit of the GPT-2 model. Each TransformerBlock consists of the following two parts in sequence: (1) Multi-head self-attention layer: This layer first processes the input vector The mapping is represented as a matrix of query Q, key K, and value V, and attention weights are calculated. ,in For Q and K vector dimensions, softmax transforms the similarity into a probability distribution; (2) Feedforward fully connected network: performs two layers of linear transformation and nonlinear activation on the vector at each position; finally, the output is processed by residual connection and layer normalization to obtain the hidden state of this layer. ; 2) Expansion of the thesaurus in the field of fiber processing; Candidate terms were selected from the fiber process training dataset. These candidate terms were specialized terms covering the entire fiber manufacturing process. A fiber domain saliency score was calculated for each candidate term. Where t represents any candidate word. The TF-IDF value of word t on the fiber technology training dataset is used to measure the importance of t in the fiber data. The TF-IDF value of word t in ordinary daily text is used to measure the frequency of t in daily corpora. IDF For inverse document word frequency, , ; Judgment: If the fiber domain saliency score of any candidate term is greater than or equal to the set threshold, then the candidate term is determined to be a new fiber core term. The N newly added fiber core terms obtained through the determination are merged into the original GPT-2 vocabulary to form a hybrid vocabulary; 3) Initialize the word vectors of newly added terms; The word vectors of the N newly added fiber domain terms selected in step 2) are initialized using a semantic nearest neighbor transfer and noise injection strategy. The specific method is as follows: Calculate the semantic similarity between each newly added fiber domain term t and all words in the original vocabulary; based on semantic similarity, select the top n nearest neighbors with the highest similarity to t from the original vocabulary. ; Calculate the arithmetic mean of the nearest neighbor word vectors: ,in The original GPT-2 embedding matrix is ​​used; Gaussian noise is superimposed to generate word vectors for the newly added fiber domain term t: in, It is superimposed Gaussian noise; The word vectors corresponding to the N newly added fiber core terms are merged with the original GPT-2 embedding matrix to form a new embedding matrix; 4) Process-aware location coding; Inject process timing features into the model's location encoding layer and encode process stage identifiers as additional location bias terms; Where pos is the absolute position number of the word in the sequence, i is the dimension sharding index, and d is the hidden layer dimension; These are the stage feature weight coefficients. For learnable projection matrices, This is a process stage identifier vector, using static encoding. ; The training of the pre-defined fiber technology large language model using the fiber technology training dataset employs a hierarchical, progressive fine-tuning training method, specifically including: Phase 1: Fine-tuning of newly added word vectors; In this phase, the base model is frozen, and the original parameters of GPT-2 remain unchanged. The semantic embedding of the newly added fiber domain terms is ensured to be stable by training the word vector parameters of the newly added fiber domain terms. Phase Two: Unfreezing of High-Level Transformers; In this stage, the last four Transformer layers of the GPT-2 model are unfrozen, and word vectors of newly added fiber domain terms are trained to adjust the model's ability to adjust high-level semantic composition. Phase 3: Fine-tuning of the entire model; This stage involves unfreezing all model parameters and performing global optimization.

2. The method according to claim 1, characterized in that, The fiber process coding rules also include anomaly handling principles; The exception handling principles include: When data is missing, it is marked as NA; When contradictory data is found, retain the original data and mark it as ERR.

3. The method according to claim 1, characterized in that, Also includes: The raw data of the fiber process are processed using a large language model; During processing, preset prompts and the original data are input into the large language model. The large language model decomposes the fiber preparation process contained in the original data into component modules, process modules, structural modules, and performance and functional modules according to the fiber process coding rules, thereby generating training samples. All training samples constitute the fiber process training dataset.

4. The method according to claim 1, characterized in that, Also includes: The fiber process codes output by the trained fiber process big language model are dynamically verified. The dynamic verification process includes checking for spelling errors, parameter errors, and format inconsistencies in the process flow in the fiber process code, and then converting the fiber process code into a natural language description.

5. A fiber processing intelligent generation device based on a large language model, characterized in that, include: The fiber process coding rule design module is used to design standardized and parsable fiber process coding rules. The training dataset construction module is used to collect raw fiber process data and perform structured splitting of the raw data according to the fiber process coding rules to obtain a fiber process training dataset. The fiber process large language model training module is used to train a preset fiber process large language model using the fiber process training dataset. The fiber process coding generation module is used to automatically generate fiber process codes that meet user requirements using the trained fiber process large language model. The fiber process coding rules include: 1) The composition module describes the raw material formula required for fiber production; 2) Process module, used to standardize the description of each processing step in fiber production, and to clarify the process type, parameter values, execution conditions and scope of each step; 3) Structural module, used to record the microscopic and macroscopic structural features of the finished fiber product to verify that the structure of the process product meets expectations; 4) Performance and Function Module: This module defines the various performance indicators and functional characteristics of fiber materials, specifies the test standards, conditions, and effective range, and evaluates the application value of the processed products. The construction process of the fiber process large language model is as follows: 1) The GPT-2 model is used as the base model; The structure of GPT-2 is as follows: in, This is the initial hidden layer. This is the lth hidden layer in the middle. For the final number of attention heads, For the input word ID matrix, For word embedding matrix, For the position encoding matrix, For word probability distribution; TransformerBlock represents the basic computational unit of the GPT-2 model. Each TransformerBlock consists of the following two parts in sequence: (1) Multi-head self-attention layer: This layer first processes the input vector The mapping is represented as a matrix of query Q, key K, and value V, and attention weights are calculated. ,in For Q and K vector dimensions, softmax transforms the similarity into a probability distribution; (2) Feedforward fully connected network: performs two layers of linear transformation and nonlinear activation on the vector at each position; finally, the output is processed by residual connection and layer normalization to obtain the hidden state of this layer. ; 2) Expansion of the thesaurus in the field of fiber processing; Candidate terms were selected from the fiber process training dataset. These candidate terms were specialized terms covering the entire fiber manufacturing process. A fiber domain saliency score was calculated for each candidate term. Where t represents any candidate word. The TF-IDF value of word t on the fiber technology training dataset is used to measure the importance of t in the fiber data. The TF-IDF value of word t in ordinary daily text is used to measure the frequency of t in daily corpora. IDF For inverse document word frequency, , ; Judgment: If the fiber domain saliency score of any candidate term is greater than or equal to the set threshold, then the candidate term is determined to be a new fiber core term. The N newly added fiber core terms obtained through the determination are merged into the original GPT-2 vocabulary to form a hybrid vocabulary; 3) Initialize the word vectors of newly added terms; The word vectors of the N newly added fiber domain terms selected in step 2) are initialized using a semantic nearest neighbor transfer and noise injection strategy. The specific method is as follows: Calculate the semantic similarity between each newly added fiber domain term t and all words in the original vocabulary; based on semantic similarity, select the top n nearest neighbors with the highest similarity to t from the original vocabulary. ; Calculate the arithmetic mean of the nearest neighbor word vectors: ,in The original GPT-2 embedding matrix is ​​used; Gaussian noise is superimposed to generate word vectors for the newly added fiber domain term t: in, It is superimposed Gaussian noise; The word vectors corresponding to the N newly added fiber core terms are merged with the original GPT-2 embedding matrix to form a new embedding matrix; 4) Process-aware location coding; Inject process timing features into the model's location encoding layer and encode process stage identifiers as additional location bias terms; Where pos is the absolute position number of the word in the sequence, i is the dimension sharding index, and d is the hidden layer dimension; These are the stage feature weight coefficients. For learnable projection matrices, This is a process stage identifier vector, using static encoding. ; The training of the pre-defined fiber technology large language model using the fiber technology training dataset employs a hierarchical, progressive fine-tuning training method, specifically including: Phase 1: Fine-tuning of newly added word vectors; In this phase, the base model is frozen, and the original parameters of GPT-2 remain unchanged. The semantic embedding of the newly added fiber domain terms is ensured to be stable by training the word vector parameters of the newly added fiber domain terms. Phase Two: Unfreezing of High-Level Transformers; In this stage, the last four Transformer layers of the GPT-2 model are unfrozen, and word vectors of newly added fiber domain terms are trained to adjust the model's ability to adjust high-level semantic composition. Phase 3: Fine-tuning of the entire model; This stage involves unfreezing all model parameters and performing global optimization.

6. An electronic device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method according to any one of claims 1-4.

Images