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

Abstract

The invention 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 a standardized and analyzable fiber process coding rule; collecting original data of a fiber process, and performing structural splitting on the original data according to the fiber process coding rule to obtain a fiber process training data set; training a preset fiber process big language model by using the fiber process training data set; and automatically generating a fiber process code meeting user requirements by using the trained fiber process large language model. According to the method, the bottleneck that traditional fiber production depends on artificial experience and is insufficient in standardization is broken through, automatic generation of a process scheme is achieved, and the fiber manufacturing industry is promoted to be transformed to data-driven precision and intelligent transformation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of fiber material process design, and particularly relates to a fiber process intelligent generation method and system based on a large language model, which is suitable for process design scenarios of various fiber materials such as carbon fibers and high polymer synthetic fibers. BACKGROUND

[0002] Fiber material production process design has long been faced with the dual challenges of knowledge management and technological innovation. Traditional methods rely on expert experience to build process systems, which have three defects: (1) risk of knowledge transmission gap, it is difficult for senior engineers to effectively convert their tacit knowledge into reusable process rules; (2) insufficient standardization of process files, different enterprises use different parameter systems and process descriptions, resulting in high failure rate of cross-platform process transplantation; (3) long research and development cycle, new process development needs to go through a linear process of literature research, scheme design and trial verification, which takes a long time.

[0003] The existing method and system for assisting in designing new materials by using a large language model (CN118824432A) uses a large language model to encode and map material composition and process parameters to generate a performance prediction model to assist in the design of new materials, but its core focuses on the construction of a material performance prediction model, and has not yet solved the problem of obtaining and preprocessing high-quality fiber material "composition-process-structure-performance / function" whole-chain preparation data, and lacks the ability to model the logical constraints of the fiber production multi-process, making it difficult to automatically generate a whole-process process scheme.

[0004] In summary, large language models have shown strong cross-modal data processing capabilities in the material genome project, but have not yet been effectively applied in the field of fiber process design. Existing AI models mostly focus on molecular structure prediction and lack the ability to model process chain logical constraints, resulting in basic errors such as parameter conflicts and process inversion in the generated process scheme, which seriously hinders the practicalization process. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art and provide a fiber process intelligent generation method and system based on a large language model. The present application breaks through the bottleneck of traditional fiber production relying on manual experience and lack of standardization, realizes the automatic generation of process schemes, and promotes the transformation of the fiber manufacturing industry to data-driven precision and intelligence.

[0006] The first aspect of the present application provides a fiber process intelligent generation method based on a large language model, comprising:

[0007] designing standardized and analyzable fiber process coding rules;

[0008] Collecting fiber process raw data, structurally splitting the raw data according to the fiber process coding rules to obtain a fiber process training data set;

[0009] Training a preset fiber process large language model using the fiber process training data set;

[0010] Generating a fiber process code meeting user needs automatically using the trained fiber process large language model.

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

[0012] 1) a component module for describing the raw material formula required for fiber production;

[0013] 2) a process module for standardizing the description of each processing step of fiber production, and clearly defining the process type, parameter value, execution condition and range of each step;

[0014] 3) a structure module for recording the microstructure and macrostructure characteristics of the fiber product to verify that the structure of the process product meets the expectations;

[0015] 4) a performance and function module for defining various performance indicators and functional characteristics of the fiber material, and indicating the test standards, conditions and effective range to evaluate the application value of the process product.

[0016] In one embodiment of the present application, the fiber process coding rules further include an exception handling principle;

[0017] The exception handling principle includes:

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

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

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

[0021] Processing the fiber process raw data using a large language model; when processing, inputting a preset prompt word and the raw data into the large language model, the large language model disassembles the fiber preparation process contained in the raw data into a component module, a process module, a structure module, and a performance and function module according to the fiber process coding rules, thereby generating training samples; all training samples constitute a fiber process training data set.

[0022] In one embodiment of the present application, the construction process of the fiber process large language model is as follows:

[0023] 1) using GPT-2 model as a base model;

[0024] where GPT-2 is structured as follows:

[0025]

[0026]

[0027]

[0028] where, is the initial hidden layer, is the intermediate l-th hidden layer, is the number of final attention heads, is the input word ID matrix, is the word embedding matrix, is the position encoding matrix, is the word probability distribution;

[0029] TransformerBlock represents the basic computational unit of the GPT-2 model, and each TransformerBlock is composed of the following two parts in sequence: (1) multi-head self-attention layer: this layer first maps the input vector to query Q, key K and value V matrices, and calculates the attention weight , where is the Q, K vector dimension, and softmax converts the similarity into a probability distribution; (2) feedforward fully connected network: two layers of linear transformation and nonlinear activation are performed on the vector of each position; finally, the output is connected by residual connection and layer normalization to obtain the hidden state of this layer .

[0030] 2) Fiber process field vocabulary expansion;

[0031] Filter the candidate vocabulary from the fiber process training data set, and the candidate vocabulary is a professional term covering the whole process of fiber manufacturing; calculate the fiber field significance score Score of the candidate vocabulary:

[0032]

[0033] where t is any candidate vocabulary, represents the TF-IDF value of word t on the fiber process training data set, to measure the importance of t in the fiber data; represents the TF-IDF value of word t in ordinary daily text, to measure the common degree of t in the daily corpus; , IDF is the inverse document word frequency, , ;

[0034] Determination: if the fiber field significance score Score of any candidate word is greater than or equal to the set threshold, the candidate word is determined to be a new fiber core term;

[0035] The N new fiber core terms obtained by determination are merged into the original GPT-2 word table to form a mixed word table;

[0036] 3) New term word vector initialization;

[0037] The word vector of the N new fiber field terms screened out in step 2) is initialized using the semantic neighbor migration and noise injection strategy, and the specific method is as follows:

[0038] Calculate the semantic similarity between each new fiber field term t and all words in the original word table; based on the semantic similarity, select the top n neighbor words with the highest similarity to t from the original word table ; Calculate the arithmetic mean of the neighbor word vectors: , where is the GPT-2 original embedding matrix; superimpose the Gaussian noise to generate the word vector of the new fiber field term t:

[0039]

[0040] where, is the superimposed Gaussian noise ;

[0041] Merge the word vectors corresponding to the N new fiber core terms with the GPT-2 original embedding matrix to form a new embedding matrix;

[0042] 4) Process-aware position encoding;

[0043] Inject process timing features in the position encoding layer of the model, and encode the process stage identifier as an additional position bias term;

[0044]

[0045]

[0046] where pos is the absolute position sequence number of the word in the sequence, i is the dimension slice index, and d is the hidden layer dimension; is the stage feature weight coefficient, is the learnable projection matrix, is the process stage identifier vector, which is statically encoded, .

[0047] In a specific embodiment of the present application, the training of the preset fiber process large language model using the fiber process training data set adopts a hierarchical progressive fine-tuning training manner, specifically including:

[0048] Phase one: new word vector fine-tuning;

[0049] In this phase, the base model is frozen, and the original parameters of GPT-2 are kept unchanged; the word vector parameters of the newly added fiber field terms are trained to ensure the stability of the semantic embedding of the newly added fiber field terms;

[0050] Phase two: high-level Transformer unfreezing;

[0051] In this phase, the last 4 layers of the GPT-2 model are unfrozen and the word vectors of the newly added fiber field terms are trained to adjust the high-level semantic combination ability of the model;

[0052] Phase three: whole model fine-tuning;

[0053] In this phase, all parameters of the model are unfrozen for global optimization.

[0054] In a specific embodiment of the present application, it further includes:

[0055] The fiber process code output by the trained fiber process large language model is dynamically verified;

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

[0057] The second aspect embodiment of the present application proposes a fiber process intelligent generation device based on a large language model, which includes:

[0058] A fiber process coding rule design module is used to design a standardized and analyzable fiber process coding rule;

[0059] A training data set construction module is used to collect fiber process original data, structure and split the original data according to the fiber process coding rule, and obtain a fiber process training data set;

[0060] A fiber process large language model training module is used to train a preset fiber process large language model using the fiber process training data set;

[0061] A fiber process coding generation module is used to automatically generate fiber process codes meeting user needs using the trained fiber process large language model.

[0062] The third aspect of the present application provides an electronic device, comprising:

[0063] at least one processor; and a memory connected with the at least one processor in communication;

[0064] The memory stores instructions executable by the at least one processor, and the instructions are configured to execute the above-mentioned fiber process intelligent generation method based on a large language model.

[0065] To achieve the above-mentioned embodiments, the fourth aspect of the present application provides a computer readable storage medium, which stores computer instructions for executing a fiber process intelligent generation method based on a large language model.

[0066] The characteristics and benefits of the present application are:

[0067] 1. Breakthrough traditional process design bottleneck, solve knowledge gap and standardization problem

[0068] The fiber process coding rule (FPC) invented by the present application defines the expert implicit knowledge as a machine interpretable standardized parameter system through the four module structure of "component-process-structure-property", unifies the cross-enterprise process description format, and solves the problem of transplantation failure caused by knowledge inheritance relying on manual and parameter system confusion in traditional methods. Through dynamic range parameters and nested sub-process design, it is compatible with complex process scenarios, and significantly reduces the enterprise technology reuse cost.

[0069] 2. Build a field-enhanced large model to realize precise process chain logic control.

[0070] Through field vocabulary expansion, process sensing position coding and hierarchical progressive fine-tuning technology, the present application enables the general large language model to have the ability of fiber process chain logic modeling. The model can automatically identify raw material fluctuations and dynamically adjust process parameters, while supporting the precise generation of nested processes, improving the professional degree of generated content and process stability, and solving the practicalization obstacles of existing AI model process inversion and parameter incompatibility.

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

[0072] The successful application of the present application will greatly shorten the research and development cycle of fiber process optimization and spinning parameter design, and reduce the trial verification cost. Through the bidirectional conversion mechanism of FPC coding and natural language, the process documents conforming to the industrial standards can be automatically output, reducing the manual workload, and providing a standardized process interface for the intelligent factory of the enterprise, realizing the parameter linkage optimization of multiple devices. It is expected to form a billion-level market application in the field of high-end materials such as aerospace and biomedical fibers, and promote the transformation of the industry from "experience trial and error" to "data-driven". 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], format "Main material = content: unit", e.g. "PAN = 88%".

[0087] Sub-process nesting: production process of main material is nested with identifier ... Package, sub-process can contain other modules (such as process steps).

[0088] Auxiliary material: additional materials are separated by semicolon ";" after main material, and the content of additional materials is described.

[0089] Special treatment:

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

[0091] Material flow: use arrow symbol "→" to represent addition order, for example "AgNO3 → H2O" means silver nitrate is added to water.

[0092] Additional attribute: state or condition is noted in parentheses after the material, such as "PLA (granular) = 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] Explanation:

[0102] Initiator AIBN (5%) is synthesized by polymerization reaction (60℃, nitrogen protection, 24 hours). Auxiliary material: add solvent DMSO (10%) and carbon nanotubes (CNT, 2wt%).

[0103] b) Process module (Process, Proc).

[0104] Process module is used to standardize the description of each processing step of fiber production, and to clearly define the process type, parameter value, execution condition and range of each step, to support the logical sequence modeling of complex multi-step process, and to ensure the reasonable and conflict-free process chain.

[0105] Rule description:

[0106] Process step: identified by [Proc] <step> ...< / step> Package, each step starts with [Proc-Type = process type], for example [Proc-Type = electrospinning].

[0107] Parameter definition: Format as parameter name = value: unit (condition), e.g., "Voltage = 20: kV (generator, ±5%)".

[0108] Special treatment:

[0109] Range parameter: Use a short dash "-" to indicate the parameter range, such as "Temp = 180-220: °C".

[0110] Multi-condition description: Add multiple conditions in parentheses, such as "FlowRate = 1.2: ml / h (pump, accuracy 0.1%)".

[0111] Example:

[0112] <step>

[0113] [Proc-Type=Electrospinning]

[0114] Voltage=20:kV (Generator, ±5%)

[0115] Flow Rate=1.2:ml / h (Syringe Pump)

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

[0117] < / step>

[0118] <step>

[0119] [Proc-Type: After Stretching]

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

[0121] < / step>

[0122] Interpretation:

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

[0124] c) Structure module (Struct).

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

[0126] Rule description:

[0127] Parameter definition: identified by [Struct], format as "parameter name = value: unit (method = measurement method, additional information)", e.g., "FiberDia = 500: nm (method = SEM, SD = 50)".

[0128] Special treatment:

[0129] Distribution description: If the pore size distribution is Gaussian distribution, mark "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] Interpretation:

[0137] Fiber diameter: 500 nm (SEM measurement, standard deviation 50 nm). Pore size: 30-100 nm (Gaussian distribution). Orientation: 75° (X-ray diffraction measurement). Crystallinity: 45% (Differential Scanning Calorimetry measurement).

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

[0139] Performance and functional modules are used to define the mechanical, electrical, antibacterial, UV-resistant, etc. performance indicators and functional characteristics of the fiber material, indicating the test standards, conditions and effective range, for evaluating the application value of the process product.

[0140] Rule description:

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

[0142] Functional properties: identified by [Func], format is "Property 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: °C, RH = 45%)

[0147] [Func]

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

[0149] UV-Resist = √ (range = 300-400: nm)

[0150] Interpretation:

[0151] Mechanical properties: tensile strength 120 MPa (ASTM standard test). Conductivity: 1 x 10³ S / m (25°C, relative humidity 45%). Antibacterial property: 99.9% inhibition rate against E. coli. Anti-ultraviolet: effective in the 300-400 nm waveband.

[0152] Further, the FPC coding rules also include:

[0153] Abnormal processing principle:

[0154] Data missing: when data is missing, mark as NA, for example [Perf] Conductivity = NA (test = not tested).

[0155] Contradictory data: when contradictory data occurs, keep the original data and mark ERR, for example [Struct] FiberDia = 200: nm / 250: nm (ERR, method = SEM / AFM).

[0156] Wherein, the meaning of the symbols can be described by the following table:

[0157] Table 1. Symbol meaning table in FPC rules in the method described in this embodiment

[0158] Symbol Name Application scenario Example ; Parameter separator Separate different elements within the same module / step Voltage = 20: kV; Flow Rate = 1.2: ml / h → Material flow direction symbol Indicate the order of addition (A→B means A is added to B) AgNO3→H2O () Condition specifier Add parameter supplement or limit condition Temp = 60: °C (N2) Conductivity = 1e3: S / m (25: °C) : Unit separator Separate numerical value and unit Time = 3: h Fiber Dia = 500: nm - Range marker Indicate the parameter range or gradient Temp = 180-220: °C Ag3PO4 = 10-30% = Assignment symbol Connect parameter name and numerical value

[0159] In this embodiment, the FPC coding rules convert the complex fiber production process into a machine interpretable format through modularized and standardized parameter definition, support nested processes, dynamic range parameters and exception annotations, and provide a structured data basis for subsequent AI model training and process generation.

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

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

[0162] The collected data can be structured by using common general large language model tools. Online APIs such as OpenAI GPT and Deepseek can be called. The specific method is to register an official platform account, obtain API access rights and keys, install API calling libraries and write Python processing scripts. The general large language model official page can also be directly used for dialogue processing. The specific method is to input specific prompt words and documents into the general large language model at the same time. The general large language model will output data in the corresponding format according to the requirements of the specific prompt words. A structured extraction instruction Doc2FPC-prompt prompt word template needs to be designed to support the extraction of fiber process related structure data from the literature data and convert it into a standard format that can be used for subsequent model training. In this process, the fiber preparation process in the original text will be disassembled into four modules of composition, process, structure and performance and function according to the FPC coding rules. In this embodiment, the collected original literature data content needs to include complete process chain description, such as material formula, processing steps, product structure and performance index, and the terminology needs to be standardized. The core features of the processed data include standardized parameter format, modular classification and nested hierarchical structure.

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

[0164] You are a polymer material expert and need to extract fiber process coding (FPC) from the document, strictly following the following rules:

[0165] Synonym unification: terms with the same meaning need to be expressed uniformly.

[0166] Single document and single FPC: only one complete FPC code is extracted from each document.

[0167] Only output FPC: no additional explanatory content is added, and the standardized FPC text is directly output.

[0168] Fiber process coding (FPC) is as follows: (here, the fiber process coding in step 1 is supplemented)

[0169] Through the above steps, a large amount of FPC structured data can be obtained according to the number of original documents. The format is JSON classified by four modules, covering material ratio, process parameters, fiber morphology and mechanical properties, etc. The training set size of this embodiment needs to cover at least 1,000 training samples, covering mainstream process types such as carbon fiber and nanofiber. Special scenarios (such as biomedical fibers) need to supplement additional samples to more than 2,000 to ensure the generalization ability of the model to long-tail data. At the same time, the sample proportion of different process types needs to be balanced, for example, to avoid too much spinning process data and too little heat setting data to prevent the model from producing bias.

[0170] The training sample in one embodiment of the present application is 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%); Flow Rate = 1.2: ml / h (Syringe Pump); Collector = Al-foil (Rotating, 500: rpm)< / step>

[0173] <step>[Proc-Type = post-stretching] 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: °C, RH = 45%)

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

[0177] Interpretation: The fiber process code (FPC) describes the preparation process of PAN-based material: the main material is PAN (88%), and its synthesis sub-process is realized through polymerization reaction (AN monomer accounts for 95%, initiator AIBN accounts for 5%, and the reaction is carried out at 60°C under nitrogen protection for 24 hours), and solvent DMSO (10%) and carbon nanotube (CNT, 2wt%) are added. The process module includes two stages of electrospinning and post-stretching: electrospinning uses a voltage of 20kV (±5% fluctuation, generator control), a flow rate of 1.2ml / h (syringe pump driven), and an aluminum foil collector rotating at 500rpm; the post-stretching process is set to a stretching ratio of 3.5 times and a stretching speed of 10mm / min. The resulting fiber structure characteristics 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 a crystallinity of 45% (DSC analysis). Performance tests show that the tensile strength reaches 120MPa (ASTM-D638 standard), the electrical conductivity is 1e3 S / m (25°C / 45% humidity environment), and it also has 99.9% E. coli inhibition rate and 300-400nm UV protection function.

[0178] 3) Build a fiber process large language model.

[0179] The fiber process large language model of the embodiment is based on a generative pre-trained transformer (GPT) architecture, which consists of the following parts: 1, an embedding layer, which maps input words to word vectors through an embedding matrix; 2, a position encoding layer, which adds a positional bias term to enhance the model's understanding of the process stage order based on standard absolute position encoding; 3, a transformer (Transformer) encoding layer, which is a multi-layer Transformer of GPT itself, including multi-head self-attention, residual connection, feedforward neural network, etc., for modeling sequence context and generating logic; 4, an output layer, which maps the Transformer output back to the word table dimension to generate a probability distribution, and the system outputs 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 model is used as the base model of the fiber process large language model, and the specific model size is selected according to the actual amount of JSON data obtained. The specific rules are as follows: if the sample size of the training set is between 1k-5k, select the GPT-2 Small model with a hidden layer dimension of 768 and an attention head number of 12 as the base model; if the sample size of the training set is between 5k-20k, select the GPT-2 Medium model with a hidden layer dimension of 1024 and an attention head number of 16 as the base model; if the sample size of the training set is between 20k-100k, select the GPT-2 Large model with a hidden layer dimension of 1280 and an attention head number of 20 as the base model; if the sample size of the training set is more than 100k, select the GPT-2 XL model with a hidden layer dimension of 1600 and an attention head number of 25 as the base model.

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

[0183]

[0184]

[0185]

[0186] wherein, is the initial hidden layer, is the intermediate l-th layer hidden layer, is the final number of attention heads, is the input word ID matrix, is a word embedding matrix, is a position encoding matrix, is a word probability distribution. Where TransformerBlock represents the basic computational unit of GPT-2 model, each TransformerBlock is composed of the following two parts in sequence: (1) Multi-head self-attention layer. This layer first maps the input vector to query Q, key K and value V matrices, and calculates the attention weight , where is the dimension of Q, K vectors, and softmax converts the similarity into a probability distribution. (2) Feedforward fully connected network. Two layers of linear transformation and nonlinear activation are performed on the vector of each position. Finally, the output is obtained through residual connection and layer normalization to get the hidden state of this layer.

[0187] 3-2) Fiber process field vocabulary expansion.

[0188] In this embodiment, high-frequency professional terms in the fiber process are screened through statistical analysis and domain knowledge. The specific method is:

[0189] First, the term screening is performed on the fiber process related structure data, where the candidate vocabulary covers the professional terms of the whole process of fiber manufacturing, mainly including five categories of materials, processes, parameters, structures and performances: the material category involves main materials (such as PAN, PLA), solvents (such as DMSO, DMF) and functional components (such as AgNPs, GO); the process category contains process types such as electrospinning and pre-oxidation, and device terms such as spinning head and high-voltage generator; the parameter category covers process parameters such as voltage and flow rate, and physicochemical indexes such as viscosity and molecular weight; the structure and performance category focuses on characteristics such as fiber diameter and tensile strength, and test methods such as SEM and XRD; in addition, it also includes special symbols and units such as N2 protection and rpm.

[0190] In this embodiment, the source of the candidate vocabulary is the fiber process training data set obtained in step 2), and the candidate vocabulary can be selected manually or directly used after tokenization. The fiber field significance score Score of the candidate vocabulary is calculated:

[0191]

[0192] Where t is any candidate vocabulary, represents the TF-IDF value of word t in the fiber process training data set, to measure the importance of t in the fiber data; represents the TF-IDF value of word t in the general daily text (such as general corpus), to measure the common degree of t in the daily corpus; , IDF is the inverse document word frequency, , The more times a word appears in the current training sample (the higher the TF), and the fewer times it appears in all training samples (the higher the IDF), the greater the TF-IDF of the word, and the more important the word. Therefore, if a word has a greater for the fiber process data, the smaller the for the general daily document, the greater the significance of the fiber field.

[0193] In one specific embodiment of the present application, the threshold value Score is set to be greater than or equal to 0.8. If a word has a high score in some samples and a low score in some samples, the system uses the mean value of Score(t) calculated by the training set as a whole to determine and exclude general words. Finally, the number of newly added fiber field terms is N, and the original GPT-2 vocabulary (50257 words) is combined with the newly added terms using SentencePiece to form a mixed vocabulary of (50257+N) words.

[0194] 3-3) Initialization of the word vector of the newly added term.

[0195] In this embodiment, the word vector of the N newly added fiber field terms selected in step 3-2) is initialized using the semantic neighborhood transfer and noise injection strategy, and the specific method is as follows:

[0196] The semantic similarity of each newly added fiber field term t and all words in the original GPT-2 vocabulary (50257 words) is calculated using a pre-trained Sentence-BERT model. Based on the semantic similarity of all words, the top n nearest neighbor words 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; in this embodiment, n=5. The arithmetic mean of the neighborhood word vectors is calculated: where is the GPT-2 original embedding matrix. Superimpose Gaussian noise to generate the word vector of the newly added fiber field term t:

[0197]

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

[0199] 3-4) Process-aware position encoding.

[0200] In the fiber process generation task, the stage characteristics of the process (such as raw material processing -> spinning forming -> post-processing) are more critical than the absolute position. The traditional position coding only captures the word order and cannot distinguish the position relationship within different process stages; this embodiment injects process timing characteristics into the position coding layer, encodes the process stage identifier as an additional position bias term, and enhances the model's perception ability of the process chain order.

[0201]

[0202]

[0203] where pos is the absolute position sequence number of the word in the sequence, i is the dimension slice index, d is the hidden layer dimension, and for GPT-2, it is 768, is the stage feature weight coefficient, is the learnable projection matrix, is the process stage identifier vector, which is statically encoded, The value of k is defined according to the complexity of the process.

[0204] 4) Use the fiber process training data set obtained in step 2) to perform hierarchical and 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 field word table and process perception coding is implemented in three stages of lightweight fine-tuning: stage one freezes the base model, keeps the original parameters of GPT-2 (including the original word vector, attention layer, feedforward neural network (FFN), etc.) unchanged, and only trains the new word vector; stage two unfreezes the high-level transformer, adjusts the high-level semantic combination ability, and adapts to the complex logic of the fiber process; stage three fine-tunes the whole model, globally optimizes the parameter distribution, and balances the field adaptation and general language ability.

[0206] Specifically, stage one: fine-tune the new word vector;

[0207] This stage only unfreezes the word vector parameters of newly added fiber field-related terms in the fiber field vocabulary, keeping the original parameters of GPT-2 (such as attention layers, FFN, etc.) frozen to ensure the general capabilities of the base model are not compromised. Training uses a low learning rate (lr=5e-5) and AdamW optimizer (betas=(0.9, 0.98), weight decay 0.01) with a batch size of 32 to accommodate memory constraints. The training period 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 achieving a recall rate of 95% or higher for validation set domain terms (such as "electrospinning" "PAN" keywords), and the cosine similarity of the newly added word vectors converges to a difference of less than 0.01 between adjacent rounds, ensuring stable semantic embedding of new words.

[0208] Stage Two: Unfreeze Top-4 Layers of High-Level Transformer;

[0209] This stage unfreezes the last 4 layers of the GPT-2 model (e.g., the 8th-11th layers when the total number of layers is 12) and the newly added word vectors, aiming to adjust the high-level semantic combination ability to adapt to the complex logic of fiber processes. The training parameters use a hierarchical learning rate strategy, with a learning rate of 1e-4 for high-level Transformer layers and a learning rate of 3e-5 for word vector layers. The optimizer is still AdamW (betas=(0.9, 0.999), weight decay 0.1). The batch size is dynamically increased to 64 to utilize the capacity of unfreezed parameters, and the training period is extended to 20-50 rounds with an early stop condition of no significant decrease in validation loss for 5 consecutive rounds. The termination criteria require that the process chain logic coherence (such as "aggregation -> spinning -> post-stretching" process generation accuracy exceeding 90%) be met.

[0210] Stage Three: Fine-tuning of the entire model;

[0211] The final stage unfreezes all parameters of the model (including Embedding, attention layers, FFN, etc.), performing global optimization to balance domain adaptation 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 (Dropout rate 0.2, gradient clipping threshold 1.0). The training period is extended to 50-100 rounds with an early stop mechanism set to no improvement in validation loss for 10 consecutive rounds. The termination condition requires that the domain task indicators (such as process parameter generation accuracy, F1 value) reach the pre-set threshold (such as F1≥0.92). During training, cosine annealing learning rate scheduling (period =100, minimum learning rate 1e-7) is used, and the performance of mixed validation domain tasks and general tasks is verified every round to ensure that the model is accurate in the fiber process scenario while retaining basic language understanding capabilities.

[0212] 5) Fiber process intelligent generation.

[0213] In this step, based on the fiber process large language model (FiberGPT) trained in step 4) and the standard process framework, the system will automatically generate a fiber production process that meets the user's needs. The user inputs specific performance or functional requirements, process conditions, and the characteristics of the target material, and FiberGPT will automatically generate the fiber process code corresponding to the fiber process flow, i.e. FPC. These process flows include raw material selection, process steps, structure and performance, etc. to ensure that the generated process scheme meets the actual production requirements.

[0214] In this embodiment, the user input can specify performance targets, process constraints and material characteristics, and the system will parse and generate a complete FPC code containing component selection, spinning parameters and post-processing conditions. For example: "Please generate an antibacterial nanofiber production process suitable for medical dressings, with fiber diameter controlled within 200-300 nm."

[0215] Optionally, the preliminary process code FPC generated in this embodiment can also 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, and will be converted into a natural language description, conforming to industry standards and facilitating engineers to understand and execute. The specific method is: use GPT-o1, Deepseek-R1, etc. Online API and FPC2Doc-prompt prompt words for translation and natural language optimization to automatically generate standard fiber production process documents. These documents not only describe each process step in detail, but also provide detailed information such as operating requirements and experimental conditions to ensure that every link in the production process can be strictly executed. The process flow is finally output in the form of a standardized report, which can be directly applied to production or used for subsequent process optimization and adjustment.

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

[0217] You are an expert in the field of high polymer materials. You need to convert the fiber processing code (FPC) into a natural language description of the fiber production process according to the following rules. Only output the complete fiber production process, do not include other content.

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

[0219] a fiber process coding rule design module for designing standardized and analyzable fiber process coding rules;

[0220] a training data set construction module for collecting fiber process raw data, structurally splitting the raw data according to the fiber process coding rules to obtain a fiber process training data set;

[0221] a fiber process large language model training module for training a preset fiber process large language model using the fiber process training data set;

[0222] a fiber process coding generation module for automatically generating fiber process coding meeting user needs using the trained fiber process large language model.

[0223] It should be noted that the foregoing embodiment of the fiber process intelligent generation method based on a large language model is also applicable to the fiber process intelligent generation device based on a large language model of the present embodiment, which will not be described here. According to the fiber process intelligent generation device based on a large language model of the present embodiment, standardized and analyzable fiber process coding rules are designed; fiber process raw data is collected, and the raw data is structurally split according to the fiber process coding rules to obtain a fiber process training data set; a preset fiber process large language model is trained using the fiber process training data set; and fiber process coding meeting user needs is automatically generated using the trained fiber process large language model. Thus, the bottleneck of relying on manual experience and lacking standardization in traditional fiber production can be broken through, and automatic generation of process schemes is realized, thereby promoting the data-driven precision and intelligent transformation of the fiber manufacturing industry.

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

[0225] 1) a component module for describing the raw material formula required for fiber production;

[0226] 2) a process module for standardized description of each processing step of fiber production, and clear definition of the process type, parameter value, execution condition and range of each step;

[0227] 3) a structure module for recording the microstructure and macrostructure characteristics of the fiber product to verify that the structure of the process product meets the expectation;

[0228] 4) a performance and function module for defining various performance indicators and functional characteristics of the fiber material, and indicating the test standard, condition and effective range to evaluate the application value of the process product.

[0229] In a specific embodiment of the present application, the fiber process coding rules further include an exception handling principle;

[0230] The abnormality processing principles include:

[0231] When data is missing, mark as NA;

[0232] When there is contradictory data, keep the original data and mark ERR.

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

[0234] The fiber process raw data is processed by using a large language model; when processing, a preset prompt word and the raw data are input into the large language model, the large language model disassembles the fiber preparation process contained in the raw data into component modules, process modules, structure modules, performance and function modules according to the fiber process coding rules, thereby generating training samples; all training samples constitute a fiber process training data set.

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

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

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

[0238]

[0239]

[0240]

[0241] wherein, is the initial hidden layer, is the intermediate l-th layer hidden layer, is the number of final attention heads, is the input word ID matrix, is the word embedding matrix, is the position encoding matrix, is the word probability distribution;

[0242] TransformerBlock represents the basic calculation unit of the GPT-2 model, and each TransformerBlock is sequentially composed of the following two parts: (1) multi-head self-attention layer: the layer first maps the input vector into query Q, key K and value V matrices, and calculates the attention weight , wherein Q, K are the vector dimensions, softmax converts the similarity into a probability distribution; (2) feedforward fully connected network: two layers of linear transformation and nonlinear activation are performed on the vector of each position separately; finally, the output is obtained through residual connection and layer normalization to obtain the hidden state of this layer ;

[0243] 2) Fiber process field vocabulary expansion;

[0244] Select candidate words from the fiber process training data set, and the candidate words are professional terms covering the whole process of fiber manufacturing; calculate the fiber field significance score Score of the candidate words:

[0245]

[0246] Where t is any candidate word, TF-IDF value of word t on the fiber process training data set, to measure the importance of t in the fiber data; TF-IDF value of word t in ordinary daily text, to measure the common degree of t in the daily corpus; IDF Inverse document word frequency, , ;

[0247] Determination: if the fiber field significance score Score of any candidate word is greater than or equal to the set threshold, it is determined that the candidate word is a new fiber core term;

[0248] Merge the N new fiber core terms obtained by determination into the original GPT-2 vocabulary to form a mixed vocabulary;

[0249] 3) Initialization of new term word vectors;

[0250] Initialize the word vectors of the N new fiber field terms screened out in step 2) using the semantic neighbor migration and noise injection strategy, and the specific method is as follows:

[0251] Calculate the semantic similarity between each new fiber field term t and all words in the original vocabulary; based on the semantic similarity, select the top n neighbor words with the highest similarity to t from the original vocabulary ; Calculate the arithmetic mean of the neighbor word vectors: , where is the GPT-2 original embedding matrix; superimpose Gaussian noise to generate the word vector of the new fiber field term t:

[0252]

[0253] Where, is the superimposed Gaussian noise ;

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

[0255] 4) Process-aware position encoding;

[0256] Injecting the process timing feature in the position encoding layer of the model, and encoding the process stage identifier as an additional position bias term;

[0257]

[0258]

[0259] wherein pos is the absolute position serial number of the word in the sequence, i is the dimension slice index, and d is the hidden layer dimension; is the stage feature weight coefficient, is the learnable projection matrix, is the process stage identification vector, and is statically encoded, .

[0260] In one specific embodiment of the present application, the preset fiber process large language model is trained using the fiber process training data set in a hierarchical progressive fine-tuning training manner, specifically including:

[0261] Stage one: new word vector fine-tuning;

[0262] In this stage, the base model is frozen, and the original parameters of GPT-2 are kept unchanged; the word vector parameters of the newly added fiber field terms are trained to ensure the stability of the semantic embedding of the newly added fiber field terms;

[0263] Stage two: unfreezing the high-level Transformer;

[0264] In this stage, the last 4 layers of the GPT-2 model and the word vectors of the newly added fiber field terms are unfrozen and trained to adjust the high-level semantic combination ability of the model;

[0265] Stage three: fine-tuning of the whole model;

[0266] In this stage, all parameters of the model are unfrozen for global optimization.

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

[0268] Performing dynamic verification on the fiber process code output by the trained fiber process large language model;

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

[0270] To achieve the above-embodiment, the third aspect of the present application provides an electronic device, comprising:

[0271] at least one processor; and a memory connected with the at least one processor in communication;

[0272] The memory stores instructions executable by the at least one processor, and the instructions are configured to execute the above-mentioned fiber process intelligent generation method based on a large language model.

[0273] To achieve the above-embodiment, the fourth aspect of the present application provides a computer-readable storage medium, which stores 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 of the present disclosure described above can be a computer readable signal medium or a computer readable storage medium or any combination of the above two. The computer readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples of computer readable storage medium can include, but are not limited to, electrical connections with one or more conductive wires, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above. In the present disclosure, the computer readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or component. In the present disclosure, the computer readable signal medium can include a data signal propagating in a baseband or as a carrier wave in a carrier, which carries computer readable program code. Such a propagating data signal can take many forms, including but not limited to electromagnetic signals, optical signals or any suitable combination of the above. The computer readable signal medium can also be any computer readable medium other than the computer readable storage medium, which can send, propagate or transmit programs for use by or in conjunction with an instruction execution system, device or component. The program code contained in the computer readable medium can be transmitted by any suitable medium, including but not limited to electrical wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0275] The computer readable medium described above can be included in the electronic device described above; or can exist separately from the electronic device and can be accessed via the electronic device. The computer readable medium described above bears one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method for generating a fiber process intelligent method based on a large language model according to the embodiments described above.

[0276] Computer program code for carrying out operations of the present disclosure can be written in one or more programming languages or combinations of languages including object oriented programming languages such as Java, Smalltalk, C++ as well as conventional procedural programming languages such as "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0277] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Also, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0278] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[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 of the present technology can understand that all or part of the steps carried out by the above-mentioned embodiment methods can be completed by programs instructing related hardware, and the programs can be stored in a computer readable storage medium. When the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0283] In addition, each functional unit in each embodiment of the present application can be integrated into one processing module, or each unit can exist physically alone, or two or more units can be integrated into one module. The integrated module can be realized in the form of hardware or in the form of a software functional module. The integrated module, if realized in the form of a software functional module and sold or used as an independent product, can also be stored in a computer readable storage medium.

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

2. The method according to claim 1, characterized in that, 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 testing standards, conditions, and effective range, and evaluates the application value of the processed products.

3. The method according to claim 2, 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.

4. The method according to claim 2, 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.

5. The method according to claim 4, characterized in that, 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 then added to generate word vectors for the newly added fiber domain term t: in, 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. .

6. The method according to claim 5, characterized in that, 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.

7. The method according to claim 6, 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.

8. 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.

9. 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 1-7.

10. 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 1-7.

Images