Multi-channel intelligent cooperative alloy component or process standardization method
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-03
- Publication Date
- 2026-03-24
AI Technical Summary
Existing technologies struggle to ensure data regularity and accuracy when processing complex and variable alloy compositions and process data. Traditional methods are inefficient and lack precision, and single large language models are difficult to standardize efficiently and accurately in specialized fields.
A multi-channel intelligent collaborative approach is adopted, which works in concert through three channels: rule matching, LLM independent reasoning, and context-aware enhanced reasoning. Data is routed to the appropriate channel for normalization based on its complexity, and a process normalization channel is introduced for synchronous governance. The semantic understanding capabilities and contextual knowledge enhancement features of the large language model are utilized.
It achieves fully automated processing from non-standard data to standardized, high-confidence component information, solving the problems of low efficiency and insufficient accuracy of traditional methods in complex scenarios, and supporting the needs of downstream material research and development and performance analysis.
Smart Images
Figure CN121725945A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of data processing, and more particularly to a multi-channel intelligent collaborative alloy composition or process standardization method. BACKGROUND
[0002] In current scientific research, there are many technical difficulties in data governance in the field of materials science. Among them, alloy composition, as the core basic parameter that determines the microstructure and macroscopic performance of materials, directly affects the subsequent material research and performance analysis. Due to the individual differences in scientific literature writing style, alloy composition data often presents complex and variable non-standard forms, such as reversed element order, omitted content unit, ambiguous matrix element, and inconsistent numerical format, which makes it difficult to guarantee data regularity and accuracy. Moreover, the process description corresponding to each alloy is also chaotic.
[0003] For alloy composition data, the traditional data governance method takes the rule-based method as the core, and the regular expression matching is its typical implementation form. Its core is to define a series of high-precision regular expressions or text patterns by researchers in advance, and capture the target alloy composition string through pattern matching. For example, for Ti40Zr10Cu36Pd14 of this type of continuous concatenation expression, a special pattern can be designed to identify the element symbol and the content value immediately following it, and reorganize them in the order of element-content. This method can exhibit high processing efficiency when dealing with highly uniform and little variant texts. However, in the real and open scientific literature scene, its limitations are significant: on the one hand, due to the need for a large number of rules corresponding to various non-standard forms, the scale of the rule set expands dramatically, and the subsequent maintenance cost rises exponentially; on the other hand, for data missing scenarios such as element missing and context dependence, ordinary rule matching methods lack semantic understanding and context association capabilities, and cannot effectively repair them.
[0004] For process data, most processes are mainly described in words, and traditional natural language processing methods are difficult to accurately and losslessly reconstruct the chaotic expressions into regular forms.
[0005] Large Language Model (LLM) can learn and internalize the knowledge system, grammar structure and logical reasoning pattern in human language by pre-training on massive data such as Internet text and academic literature, and promote language information processing from pattern recognition stage to complex semantic understanding stage. Unlike regular expressions that require pre-set rules, large language models can automatically identify the logical expression of alloy composition by analyzing text, and understand the core semantics of alloy composition through training or guided models, and then generate a standardized expression form. Although general large language models can solve the defects of rule inflation and low efficiency of traditional rule matching methods, in the field of material science with extremely strong professionalism, it is difficult for a single large language model to independently achieve efficient and accurate standardization in all scenarios when facing thousands of alloy composition data with different forms and incomplete context information.
[0006] Therefore, how to solve the problem of automatic governance and repair of complex alloy composition and process data is a problem that those skilled in the art need to solve. SUMMARY
[0007] In view of the above problems, the present application provides a multi-channel intelligent collaborative alloy composition or process standardization method in order to overcome the above problems or at least partially solve the above problems.
[0008] In order to achieve the above purpose, the present application adopts the following technical solutions: A multi-channel intelligent collaborative alloy composition or process standardization method, comprising the following steps: S1, obtaining the original literature directory column and the alloy composition column in the csv format data set, establishing the association mapping between the original literature directory and the alloy composition, and performing basic cleaning on each alloy composition data; S2, evaluating the processing complexity of each alloy composition data, and routing each alloy composition data to the corresponding alloy standardization channel according to the processing complexity; S3, receiving the alloy composition data transmitted by each alloy standardization channel through a plurality of standardized processing modules corresponding one by one, and performing different standardization processing schemes on the alloy composition data to obtain standardized results; S4, verifying the standardized results output by each standardized processing module, and correcting the standardized results that do not meet the requirements of the verification results.
[0009] Further, in S1, the way to establish the association mapping between the original literature directory and the alloy composition is to generate a unique hash code for each original literature directory, bind the alloy composition column data with the original literature through the hash code, and access the local original literature through the hash code.
[0010] Furthermore, the basic cleaning in S1 includes: Remove irrelevant spaces and garbled characters from the alloy composition column data, and standardize the special character format; Filter out invalid sample data where the original literature catalog column or alloy composition column is empty in each row of data.
[0011] Furthermore, in S2, there are three alloy normalization channels: rule matching processing channel, LLM independent reasoning channel, and context-aware enhanced reasoning channel. A two-layer filtering mechanism is used to evaluate the processing complexity of alloy composition data. The first layer filtering mechanism is as follows: multiple regular expressions for standard formats are predefined by the rule base engine. If the current alloy composition data matches any regular expression, it is routed to the rule matching processing channel; if the match fails, the second layer filtering mechanism is executed. The second-layer screening mechanism is as follows: a large model is used to establish a scoring mechanism for the processing complexity of alloy composition data based on three dimensions: completeness of alloy element description, format standardization, and semantic independence. Prompt words are embedded, and the large model scores the current alloy composition data under the guidance of the prompt words. Based on the total score and the scores of key dimensions, the current alloy composition data is classified as generally complex or highly complex. If the current alloy composition data is classified as generally complex, it is routed to the LLM independent inference channel; if the current alloy composition data is classified as highly complex, it is routed to the context-aware enhanced inference channel.
[0012] Furthermore, in S3, the normalization processing modules are: rule-based normalization engine, LLM independent semantic reasoning module, and context-enhanced knowledge reasoning module, which correspond one-to-one with the rule matching processing channel, LLM independent reasoning channel, and context-aware enhanced reasoning channel, respectively. Among them, the rule-based standardization engine performs standardization processing on alloy composition data according to predefined parsing rules; The LLM independent semantic reasoning module guides the large model to perform in-situ reasoning on the current alloy composition data by setting prompt words; The context-enhanced knowledge reasoning module infers the alloy composition data by accessing the context of the original bibliographic catalog associated with the current alloy composition data.
[0013] Furthermore, the rule-based standardization engine's processing of alloy composition data includes: First, the input alloy composition data is validated for element validity using a predefined element symbol library. Then, the element composition and corresponding content values are extracted using a pre-defined regular expression, and the content values are validated. If any validation fails, the current alloy composition data is marked as abnormal. If the validation passes, the matrix element is identified and placed at the beginning of the standard format. Other alloy elements are reorganized according to the numerical-element specification order, and the detected mass percentage or atomic percentage unit markings are added. If additional explanatory information is included, the explanatory information is set in additional brackets in the standard format.
[0014] Furthermore, the process by which the LLM independent semantic reasoning module infers from alloy composition data includes: The domain identity and constraint scope of the large model are set, and the alloy composition specification task is broken down into a series of sub-tasks, including composition identification, element and content matching, and sorting. A fixed output template is set, and in-situ inference is performed on all input alloy composition data. During the inference process, if a single data point is abnormal, the data point is marked as abnormal, and the conversion results of all alloy composition data are output according to the preset output template.
[0015] Furthermore, the context-enhanced knowledge reasoning module's processing of alloy composition data includes: The original literature associated with the current alloy composition data is divided into chapters to obtain multiple chapter blocks; Each chapter block is extracted to varying degrees according to preset weights; The system integrates with a large model, constructs prompt words, and embeds them into the large model according to the preset weights of the extracted content. The task of standardizing alloy composition data is broken down into multiple sub-tasks for reasoning, and the reasoning results are output according to a preset template.
[0016] Furthermore, S4 includes: All standardized results are format-validated using a pre-defined regular expression. Standardized results that fail the format validation are marked as outliers. The confidence level of the output results of the rule-based standardization engine is set to the highest, while the confidence levels of the output results of the LLM independent semantic reasoning module and the context-enhanced knowledge reasoning module are divided into three levels: high, medium, and low. The number of samples at each confidence level is counted to obtain the candidate sample set. The higher the confidence level, the fewer samples need to be counted. Abnormal samples and candidate sample sets are used as samples to be corrected.
[0017] Furthermore, in S4, the standardized results of the corrected alloy composition are used as a new rule pattern, a new instruction set, or a new set of prompt words.
[0018] Furthermore, S1 also includes retrieving process-related data columns from the CSV format dataset. These process-related data columns include: preparation process column, rolling process column, and heat treatment process column. After S4, it also includes: S5. Introduce a process standardization channel and a semantic review and correction module. Process data is sent to the semantic review and correction module through the process standardization channel. The semantic review and correction module uses the large model to execute the process review and standardization process and outputs standardized process data. The standardized process data is randomly sampled and added to the samples to be corrected in S4.
[0019] Furthermore, in S5, the semantic auditing and correction module processes process data as follows: Design a process unit template framework to achieve process standardization goals: Determine the actual process type of each process step based on process logic; Reclassify the incorrectly classified process steps; Remove duplicate processes; Retain all valid process units that match the process unit template framework.
[0020] The original process data is broken down into multiple process parameters, which are then filled into corresponding process units. Process units corresponding to parameters not mentioned in the original process data are not output. The process units are then spliced together to form standardized process data.
[0021] As can be seen from the above technical solution, compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a multi-channel collaborative architecture that integrates rule standardization, LLM semantic reasoning, and contextual knowledge enhancement. Based on the complexity and characteristics of alloy data, it automatically routes the data to different alloy standardization channels to execute different governance schemes. It integrates the determinism of rule matching, the semantic understanding capability of large language models, and the knowledge enhancement characteristics of context engineering to achieve fully automated processing from raw non-standard data to standardized, high-confidence composition information. Finally, it outputs standardized data that meets the needs of downstream material research and development, performance analysis, and other tasks, effectively solving the technical pain points of low efficiency and insufficient accuracy of traditional methods in complex scenarios.
[0022] 2. This invention adopts the approach of classification before treatment. First, the data is preliminarily analyzed and classified. Then, prompt words and instruction sets are designed according to the requirements. This enables the large model to locate more accurately when treating and repairing data, and avoids the large model directly identifying, treating or repairing samples through a long set of general prompt words, which may lead to over-inference.
[0023] 3. This invention also introduces a process standardization channel and separates the management of alloy composition and process data to achieve synchronous management of alloy composition and process steps. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0025] Figure 1 A flowchart of the multi-channel intelligent collaborative alloy composition or process standardization method provided in Embodiment 1 of the present invention; Figure 2 A detailed flowchart of the multi-channel intelligent collaborative alloy composition or process standardization method provided in Embodiment 1 of the present invention; Figure 3 This is a flowchart of the multi-channel intelligent collaborative alloy composition or process standardization method provided in Embodiment 2 of the present invention; Figure 4 This is a detailed flowchart of the multi-channel intelligent collaborative alloy composition or process standardization method provided in Embodiment 2 of the present invention. Detailed Implementation
[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] Example 1: like Figures 1-2 As shown in the figure, this invention discloses a multi-channel intelligent collaborative alloy composition or process standardization method, including the following steps: S1. Obtain the original literature catalog column and alloy composition column from the CSV format dataset, establish the association mapping between the original literature catalog and alloy composition, and perform basic cleaning on each alloy composition data. S2. Assess the processing complexity of each alloy composition data point and route each alloy composition data point to the corresponding alloy normalization channel based on the processing complexity. S3. The alloy composition data transmitted from each alloy standardization channel is received one by one through multiple standardization processing modules, and different standardization processing schemes are performed on the alloy composition data to obtain the standardization result. S4. Verify the standardized results output by each standardization processing module, and correct any standardized results that do not meet the requirements.
[0028] The steps in Example 1 will be further explained below.
[0029] S1. Data Acquisition and Preprocessing: Obtain the alloy composition dataset from a scientific literature database in the field of heavy metal materials science, read the original literature catalog columns and alloy composition columns, and perform basic cleaning. The specific cleaning method is as follows: Remove irrelevant spaces and garbled characters from the alloy composition column data, and standardize the special character format, such as converting Zr0.5 to Zr0.5; Identify text language types using natural language processing methods, and distinguish between mixed Chinese and English expressions; Filter out invalid sample data where the original bibliography column or alloy composition column is empty in each row of data; Pre-check the basic format of component data, including the capitalization of element symbols, the rationality of content values, and whether units are labeled, and mark suspected abnormal samples.
[0030] Subsequently, a bidirectional association mechanism is employed to establish a mapping between the original literature catalog and the alloy composition. Specifically, a unique hash code is generated for each original literature catalog in the original dataset. This hash code binds the alloy composition column data to the original literature, and the hash code is used to access the local original literature, enabling data-to-literacy tracing. After association, basic checks are automatically performed: each sample row is checked for misalignments such as an empty alloy composition catalog or an empty literature catalog. If such misalignments are found, they are marked as "abnormal association samples" to prevent tracing breaks in subsequent processing. After completing these basic steps, the module adds preliminary metadata tags to each sample, including a preliminary quality assessment status, ultimately forming raw data records to be processed.
[0031] S2, Data Routing and Strategy Allocation: Specifically, it includes a two-layer filtering mechanism and three alloy normalization channels; the three alloy normalization channels are: rule matching processing channel, LLM independent inference channel, and context-aware enhanced inference channel; A two-layer filtering mechanism is used to evaluate the processing complexity of alloy composition data. The first layer filtering mechanism is as follows: multiple regular expressions for standard formats are predefined by the rule base engine. If the current alloy composition data matches any regular expression, it is routed to the rule matching processing channel and the quality status in the metadata tag of the data is marked as 0. If the match fails, the second layer filtering mechanism is executed. The second-layer screening mechanism is as follows: a large model is used to establish a scoring mechanism for the processing complexity of alloy composition data based on three dimensions: completeness of alloy element representation, format standardization, and semantic independence. Hint words are embedded, and the large model scores the current alloy composition data under the guidance of the hint words. Based on the total score and the scores of key dimensions, the current alloy composition data is classified as generally complex or highly complex.
[0032] If the current alloy composition data is classified as generally complex, it is routed to the LLM independent inference channel; if the current alloy composition data is classified as highly complex, it is routed to the context-aware enhanced inference channel.
[0033] Specifically, the current alloy composition data is scored in three dimensions: completeness of alloy element description, standardization of format, and semantic independence. 0 points represent no problem, 1 point represents a slight abnormality, and 2 points represent a serious abnormality. The total score ranges from 0 to 6 points.
[0034] Next, specialized prompts were set, with the following design logic: 1) The prompts clearly state the core task of grading the complexity of alloy composition data at the outset, limiting the application scenario to the standardized pre-classification of alloy composition in materials science literature, thus preventing the large model from deviating from the task; 2) The completeness of the alloy element description, the standardization of the format, and the semantic independence are constructed into a complexity spectrum and embedded with the prompts, allowing the large model to clearly understand the judgment criteria and basis; 3) The large model is first required to parse the target data dimension by dimension, first determining whether elements are missing, then verifying whether the format is consistent, and finally confirming whether it depends on the external context before outputting the classification results and returning the routing labels to the two channels of LLM independent inference or context-enhanced inference; 4) An instruction to prohibit over-inference is added to reduce the illusion of the large model. Finally, the total score obtained by the large model inference routes the alloy composition data to different channels.
[0035] The judgment criteria are as follows: If the total score is 1 to 3 points and the element description completeness dimension is 0 points, it is considered that the alloy composition data expression form is relatively complex and its processing difficulty is judged as generally complex. For example, the format of "90wt.% TiAl+6.5wt.% Nb" is messy, "18Cr-10Ni-Ti" matrix is not placed in front, etc., which do not meet the rule matching, but the element types and content values are not missing. Based on the content of the data itself, semantic understanding and reasoning can be independently completed to generate standardized results, and then it is sent to the LLM independent reasoning channel.
[0036] If the total score is 4 or above, or the element description completeness dimension is greater than 1 point, the alloy composition data is considered highly ambiguous and judged as highly complex. For example, statements such as "Fe-based alloy containing Cr" or "composition as in Section 1.2" without clear numerical values, or omissions or missing elements, such as "the alloy contains 6.5% Nb, the remainder is TiAl" without specifying the TiAl ratio, or "Ti (balance), Al (5.5–6.75%)" lacking other key alloying elements, make it impossible to reason based on the data itself. In such cases, it is necessary to use the source literature as external knowledge assistance, and the alloy composition data and its corresponding original literature list should be sent together into the context-enhanced reasoning channel.
[0037] S3, Collaborative Governance: The standardization processing module has three parts: rule-based standardization engine, LLM independent semantic reasoning module and context-enhanced knowledge reasoning module, which correspond one-to-one with the rule matching processing channel, LLM independent reasoning channel and context-aware enhanced reasoning channel, respectively. Among them, the rule-based standardization engine performs standardization processing on alloy composition data according to predefined parsing rules; The LLM independent semantic reasoning module guides the large model to perform in-situ reasoning on the current alloy composition data by setting prompt words; The context-enhanced knowledge reasoning module infers the alloy composition data by accessing the context of the original bibliographic catalog associated with the current alloy composition data.
[0038] Specifically, the rule-based standardization engine consists of a series of regular expression pattern libraries and dedicated parsing algorithms, capable of systematically handling more than ten typical non-standard formats, including continuous concatenation, subscript expression, numerical pre-position, and proportional description formats. The processing of alloy composition data includes: First, the input alloy composition data is validated for element validity using a predefined element symbol library. Then, the elemental composition and corresponding content values are extracted using a pre-defined regular expression. The content values are then validated. If the content is greater than 100 or the total content is greater than 100, it is considered abnormal data and marked as abnormal, and the validation fails. If any validation fails, the current alloy composition data is marked as abnormal; if the validation passes, the matrix element is identified and placed at the beginning of the standard format. Other alloy elements are reorganized according to the numerical-element specification order, and the detected mass percentage or atomic percentage unit is added. If additional explanatory information is included, it is set in additional brackets in the standard format.
[0039] The LLM independent semantic reasoning module is designed for alloy composition descriptions that cannot be directly inferred through simple rule matching but can be derived directly by the large language model using its own knowledge. This module is built upon a pre-trained large language model; in this embodiment, the aforementioned large model is used as the reasoning engine. The model is guided to complete the standardization transformation task through prompt word engineering. The specific reasoning process includes: The domain identity of the large model is set as a specialist in alloy composition standardization in the field of materials science, and a constraint is set, namely, in-situ inference based on the currently input data.
[0040] The task was then broken down into a more detailed breakdown, with the alloy composition specification task being divided into a series of sub-tasks such as composition identification, element and content matching, and sorting. A fixed output template was set, and the model was allowed to set its own confidence level, which was divided into three levels: high, medium, and low. The word segmenter was then called to encode the constructed dialogue text and generate an input tensor that the model could process.
[0041] Next, in-situ inference is performed on all input alloy composition data. During the inference process, key parameters in the large model inference generation process are fixed to force the model to produce consistent outputs for the same input. The main parameter fixed is the temperature parameter, which is set to 1.0. Sampling is disabled to reduce the randomness of the generated vocabulary, ensuring consistent outputs for the same input and thus guaranteeing the reproducibility of the treatment results. During model generation, the maximum number of new tokens is set to 512 to provide ample output space for complex composition inference. After generation, the output sequence is decoded, special tokens are removed, and pure, standardized results are extracted.
[0042] This module integrates an exception handling mechanism. If a single data entry malfunctions, it captures detailed error information and marks the data entry as a conversion failure, ensuring that the overall processing flow is not interrupted by individual exceptions. All conversion results are recorded and appended to the original dataset, forming a complete treatment trajectory. Finally, the conversion results of all alloy composition data are output according to the preset output template.
[0043] The context-enhanced knowledge reasoning module addresses data with highly complex, ambiguous, or missing elemental expressions in alloy composition, heavily reliant on the background of the original literature. This module constructs a complete literature context processing pipeline, its core being the establishment of a deep association between alloy composition data and its source literature. The specific processing steps include: The original literature associated with the current alloy composition data is segmented into chapters. The structured content of the literature is automatically identified and extracted through a predefined chapter keyword library (abstract, introduction, experimental methods, results discussion, conclusions, etc.) to obtain multiple chapter blocks.
[0044] The content of each chapter block is truncated to different degrees according to preset weights. Specifically, for each document, an adaptive content truncation strategy is adopted, and each chapter block is assigned weight according to the importance of the block. For example, core chapters such as abstract and experimental methods are assigned more than 90% weight, indicating that the context block is retained in its entirety. Auxiliary chapters such as introduction are assigned 50% weight, allowing for more content to be truncated. This ensures that the content length of each chapter meets the input limit of the large model while retaining key information.
[0045] In the context building phase, enhanced cue words containing multi-level background knowledge are constructed based on the original alloy composition data and extracted literature chapter block information. The cue word construction logic is similar to that of the LLM independent reasoning module: First, the model is defined as an expert in alloy composition analysis and standardization in the field of materials science, clarifying its task; second, the original alloy composition to be standardized is embedded to lock the processing target, and then embedded into the context system in order of weight according to the content truncation strategy; next, the complex standardization task is structurally decomposed into executable sub-tasks, covering composition identification, format standardization, element sorting, etc.; finally, the standardization target template is specified, and the model is required to uniformly return the standardization results and confidence levels (high, medium, low). The cue word setting adopts a priority design, prioritizing the integration of key chapters containing substantial composition information such as experimental methods, results, and abstracts to ensure that the model can obtain the most relevant background cues.
[0046] During batch processing, the context-enhanced knowledge reasoning module automatically retrieves relevant literature for each complex alloy composition data point in the data governance pipeline and executes context-enhanced reasoning in parallel. Stable API call parameters are used, with an optimal temperature value of 1.0 set to ensure output consistency, and sufficient token space is configured to accommodate complex reasoning processes. Furthermore, this module integrates a robust error handling mechanism; when processing a single data point fails, detailed error information is recorded, and subsequent samples are processed, ensuring the robustness of batch operations. All governance results include standardized components, confidence assessments, and analytical explanations, providing users with transparent decision-making support.
[0047] S4. Verify the standardized results output by each standardization processing module, and correct any standardized results that do not meet the requirements.
[0048] This step performs unified verification and confidence assessment on the standardized results output by each normalization module to ensure the accuracy, consistency, and usability of the output data.
[0049] Since rule-based matching governance is based on deterministic rules and has the strongest robustness among the three governance schemes, the confidence level of the rule-based standardization engine is set to the highest. The confidence levels of the LLM independent semantic reasoning module and the context-enhanced knowledge reasoning module are generated by the large model and are divided into three levels: high, medium, and low. In the subsequent unified verification process, the first layer of format verification is performed using rule matching. The regular expression of the normalized target format, namely "matrix element - content 1 element 1 - content 2 element 2 - ... - content n element n (wt.% or at.%)", is directly used to match all normalized alloy composition data. Alloy compositions that cannot be covered represent formats that fail verification, are marked and retained for subsequent correction. The second layer of content verification is performed based on the rule matching results. The regular expression of the target format is modified to extract each element in the sample and its corresponding content coefficient. All elements that have appeared are listed separately, and the content coefficients are stored in the corresponding columns. Then, based on the aforementioned four confidence levels (highest, high, medium, and low), the number of samples included in each confidence level is counted. Weights are set from low to high, i.e., the number of samples n that need to be manually corrected is specified. The highest confidence level is set to 5%, the high confidence level to 15%, the medium confidence level to 30%, and the low confidence level to 50%. Then, the specific number of samples to be taken for each confidence level is calculated, and stratified random sampling is performed to form the set to be corrected.
[0050] To achieve feedback and a closed loop, human expert knowledge is introduced to correct the standardization results. The correction targets are the aforementioned data that failed format validation and the set to be corrected. For the former, human experts trace the causes by interpreting specific format issues, while for the latter, human experts evaluate them based on format rationality, content rationality, and other aspects. The corrected standardization results are used as new rule patterns to eliminate inefficient or negatively impactful regular expressions in the rule base, or the corrected results are used as new instruction sets or new prompt word sets to optimize prompt word settings. Simultaneously, successful and unsuccessful cases of the context-enhanced knowledge reasoning module are analyzed to optimize the retrieval weight of document chapters, and the context matching strategy is adjusted based on expert feedback.
[0051] Example 2: This embodiment is similar to Embodiment 1, and the overall implementation process is as follows: Figures 3-4 As shown, the difference is that in S1, the process data columns in the CSV format dataset are also obtained. The process data columns include: preparation process column, rolling process column, and heat treatment process column.
[0052] Following S4, it also includes: S5. Introduce a process standardization channel and a semantic review and correction module. Process data is sent to the semantic review and correction module through the process standardization channel. The semantic review and correction module uses the large model to execute the process review and standardization process and outputs standardized process data. The standardized process data is randomly sampled and added to the samples to be corrected in S4.
[0053] Specifically, the three types of process data are diverted into the process standardization channel, and a prompt word template framework for the target process unit of process standardization is designed. The prompt word template design method is as follows: based on the knowledge of the heavy metal materials field, process parameter names are set as process units for the three types of processes, waiting for matching and filling in the process parameters decomposed from the original data.
[0054] Next, a large-scale semantic review and correction task will be performed: Analyze each process step in the input text one by one, determine its true process type according to the material industry logic, adjust the incorrectly classified steps to the correct process type; delete process units that appear repeatedly across processes or within the same process, and retain all valid process units that match the template.
[0055] Based on the process units in the pre-designed prompt word template, the large model is used to extract and match the process parameters in the three types of process data to form the structure of process unit: process parameter. For process parameters that are not mentioned, no output is given. The successfully matched structures are spliced together to form a standardized process unit.
[0056] After the process data is standardized, random samples are taken from it and added to the samples to be corrected in S4. Human experts then evaluate the data based on the rationality of the format, the rationality of the content, and the correctness of the process parameters.
[0057] The methods provided in the embodiments of the present invention have been described in detail above, and the description of their application scenarios is used to help understand the core ideas and implementation process of the present invention. Those skilled in the art will readily recognize that, without departing from the spirit and principles disclosed herein, adaptive modifications can be made to the element symbol library, rule matching mode, and prompt word template of the present invention to apply the present invention to other element composition standardization scenarios such as catalyst formulations and semiconductor doping composition ratios. Therefore, the embodiments in the specification should not be considered as limitations on the scope of protection of the present invention, and the scope of protection of the present invention should be defined by the appended claims.
[0058] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0059] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A multi-channel intelligent collaborative method for standardizing alloy composition or processes, characterized in that, Includes the following steps: S1. Obtain the original literature catalog column and alloy composition column from the CSV format dataset, establish the association mapping between the original literature catalog and alloy composition, and perform basic cleaning on each alloy composition data. S2. Assess the processing complexity of each alloy composition data point and route each alloy composition data point to the corresponding alloy normalization channel based on the processing complexity. S3. The alloy composition data transmitted from each alloy standardization channel is received one by one through multiple standardization processing modules, and different standardization processing schemes are performed on the alloy composition data to obtain the standardization result. S4. Verify the standardized results output by each standardization processing module, and correct any standardized results that do not meet the requirements.
2. The method as described in claim 1, characterized in that, In S1, the method for establishing the association mapping between the original literature catalog and the alloy composition is as follows: a unique hash code is generated for each original literature catalog, the alloy composition column data is bound to the original literature through the hash code, and the local original literature is accessed through the hash code.
3. The method as described in claim 1, characterized in that, The basic cleaning in S1 includes: Remove irrelevant spaces and garbled characters from the alloy composition column data, and standardize the special character format; Filter out invalid sample data where the original literature catalog column or alloy composition column is empty in each row of data.
4. The method as described in claim 1, characterized in that, In S2, there are three alloy normalization channels: rule matching processing channel, LLM independent reasoning channel, and context-aware enhanced reasoning channel. A two-layer filtering mechanism is used to evaluate the processing complexity of alloy composition data. The first layer filtering mechanism is as follows: multiple regular expressions for standard formats are predefined by the rule base engine. If the current alloy composition data matches any regular expression, it is routed to the rule matching processing channel; if the match fails, the second layer filtering mechanism is executed. The second-layer screening mechanism is as follows: a large model is used to establish a scoring mechanism for the processing complexity of alloy composition data based on three dimensions: completeness of alloy element description, format standardization, and semantic independence. Prompt words are embedded, and the large model scores the current alloy composition data under the guidance of the prompt words. Based on the total score and the scores of key dimensions, the current alloy composition data is classified as generally complex or highly complex. If the current alloy composition data is classified as generally complex, it is routed to the LLM independent inference channel; if the current alloy composition data is classified as highly complex, it is routed to the context-aware enhanced inference channel.
5. The method as described in claim 4, characterized in that, In S3, the normalization processing modules are: rule-based normalization engine, LLM independent semantic reasoning module and context-enhanced knowledge reasoning module, which correspond one-to-one with rule matching processing channel, LLM independent reasoning channel and context-aware enhanced reasoning channel, respectively. Among them, the rule-based standardization engine performs standardization processing on alloy composition data according to predefined parsing rules; The LLM independent semantic reasoning module guides the large model to perform in-situ reasoning on the current alloy composition data by setting prompt words; The context-enhanced knowledge reasoning module infers the alloy composition data by accessing the context of the original bibliographic catalog associated with the current alloy composition data.
6. The method as described in claim 5, characterized in that, The process of processing alloy composition data by the rule-based standardization engine includes: First, the input alloy composition data is validated for element validity using a predefined element symbol library. Then, the element composition and corresponding content values are extracted using a pre-defined regular expression, and the content values are validated. If any validation fails, the current alloy composition data is marked as abnormal. If the validation passes, the matrix element is identified and placed at the beginning of the standard format. Other alloy elements are reorganized according to the numerical-element specification order, and the detected mass percentage or atomic percentage unit markings are added. If additional explanatory information is included, the explanatory information is set in additional brackets in the standard format.
7. The method as described in claim 5, characterized in that, The LLM independent semantic reasoning module's reasoning process for alloy composition data includes: The domain identity and constraint scope of the large model are set, and the alloy composition specification task is broken down into a series of sub-tasks, including composition identification, element and content matching, and sorting. A fixed output template is set, and in-situ inference is performed on all input alloy composition data. During the inference process, if a single data point is abnormal, the data point is marked as abnormal, and the conversion results of all alloy composition data are output according to the preset output template.
8. The method as described in claim 5, characterized in that, The context-enhanced knowledge reasoning module processes alloy composition data in the following ways: The original literature associated with the current alloy composition data is divided into chapters to obtain multiple chapter blocks; Each chapter block is extracted to varying degrees according to preset weights; The system integrates with a large model, constructs prompt words, and embeds them into the large model according to the preset weights of the extracted content. The task of standardizing alloy composition data is broken down into multiple sub-tasks for reasoning, and the reasoning results are output according to a preset template.
9. The method as described in claim 5, characterized in that, S4 includes: All standardized results are format-validated using a pre-defined regular expression. Standardized results that fail the format validation are marked as outliers. The confidence level of the output results of the rule-based standardization engine is set to the highest, while the confidence levels of the output results of the LLM independent semantic reasoning module and the context-enhanced knowledge reasoning module are divided into three levels: high, medium, and low. The number of samples at each confidence level is counted to obtain the candidate sample set. The higher the confidence level, the fewer samples need to be counted. Abnormal samples and candidate sample sets are used as samples to be corrected.
10. The method as described in claim 9, characterized in that, S1 also includes: obtaining the process data columns from the CSV format dataset, which include: preparation process column, rolling process column, and heat treatment process column; S4 is followed by: S5. Introduce a process standardization channel and a semantic review and correction module. Process data is sent to the semantic review and correction module through the process standardization channel. The semantic review and correction module uses the large model to execute the process review and standardization process and outputs standardized process data. The standardized process data is randomly sampled and added to the samples to be corrected in S4.