Textile high-carbon emission process root cause analysis method driven by large language model

Abstract

The invention discloses a big language model-driven spinning high-carbon emission process root cause analysis method, and belongs to the technical field of carbon emission root cause analysis in industrial production. The method comprises the following steps that firstly, in order to meet the input requirement of a large model, a production process data structure and a semantic text representation method are predefined, and production process log data, graph data and the like related to domain knowledge stored in a local information management system are subjected to process summarization and textualization representation; secondly, similar case matching is carried out by using a proximity retrieval method based on production process data to form a support basis for subsequent large model analysis; and finally, carrying out two-stage confidence evaluation based on a proximity retrieval case and a large model root cause analysis result, and outputting a decision result. Compared with the existing design, the method can carry out traceability analysis and process optimization assistance on the carbon emission problem, and is beneficial to discovering the carbon emission influence factors of each specific process, thereby improving the production process.

Description

Technical Field

[0001] The present invention relates to the technical field of root cause analysis of carbon emissions in industrial production, and in particular to a root cause analysis method for a textile high-carbon emission process driven by a large language model. Background Art

[0002] Carbon emissions from production activities have had a wide and far-reaching impact on the global environment, ecosystems, and economic and social development. With the intensification of global warming, carbon emissions have become a key factor leading to the greenhouse effect and frequent extreme weather events. Large-scale carbon emissions from industrial production are one of the main factors leading to current problems. Therefore, the root cause analysis method for high-carbon emission processes in current industrial production processes has important research significance. With the development of technology, traditional manufacturing has begun to transform into industrial knowledge-intensive production. Among them, industrial knowledge-intensive production involves complex technologies, equipment, and procedures, and advocates automation and intelligence of production processes. Industrial knowledge-intensive production is a trend of industrial transformation. It is large-scale and has diverse production processes, requiring advanced technology and management. Industrial decision-making assistance can help production companies better grasp the data and information in the production process, thereby making the production process more intelligent and efficient. Root cause analysis of industrial production status is a method of industrial process mining and an important part of industrial decision-making assistance. It can help identify and solve problems in production process technology, process or energy optimization, and therefore becomes a key step in improving production efficiency and quality. In particular, decision-making assistance for high-carbon emission processes in industrial production is of great significance to promoting the process industry to develop towards low-carbon practices and transformation. This paper focuses on the root cause analysis of high carbon emissions in the context of textile production, aiming to support the goal of industrial process mining in promoting the low-carbon transformation and innovative development of the textile industry.

[0003] Traditional methods rely on expert knowledge and intervention. It is time-consuming, inefficient and can only analyze limited types of problems for the root cause analysis of high carbon emissions involving a large number of production resources in complex production processes. Large models and autonomous agent technology provide a possible solution to the above problems, improving the level of intelligent automation and evaluation credibility of auxiliary decision-making. However, the current production field faces a significant challenge in carbon emission management: there is a lack of high-precision root cause analysis methods for high-carbon emission processes in large-scale production processes. Existing carbon emission analysis technologies mainly focus on macro-level estimation of the comprehensive carbon emissions of the entire factory or park.

[0004] For example, the carbon emission analysis method for chemical parks disclosed in China's invention patent CN116109463A covers multiple modules including carbon emission source management, data collection, carbon emission calculation, carbon emission footprint analysis, and carbon emission control strategy generation. This solution aims to help enterprises fully understand the origin, quantitative data and causes of carbon emissions in chemical parks, and then generate analysis reports to provide data support for the low-carbon development of the park.

[0005] In addition, China's invention patent CN117557220A proposed a carbon emission analysis and management system based on dual-carbon multi-source data. The system integrates functional modules such as data chart analysis, data upload, carbon emission flow analysis, knowledge graph update, and carbon emission situation awareness. This method conducts detailed carbon emission analysis for high-energy-consuming industries and key energy-consuming enterprises through the effective integration and analysis of multi-source data, and presents it in the form of a visual report, thereby helping enterprises monitor overall indicators and promote their green transformation process. Although the application of these technologies provides important tools for carbon emission management, it still lacks root cause analysis for specific processes.

[0006] The above methods usually consider the textile production process as a whole, while ignoring the independent contribution of each specific process to carbon emissions during the production and processing of the product. The carbon emission factors involved in each process are different and the improvement methods are different. Although the overall carbon emission analysis method can provide a certain reference, it cannot accurately reflect the carbon emissions of each link in the production process, making it difficult to customize the optimization measures for each process. Therefore, carbon emission root cause analysis for independent specific high-carbon emission processes is the only way for current industrial production to achieve low-carbon transformation.

[0007] In order to solve the above problems, the present invention proposes a root cause analysis method for high carbon emission processes in the textile industry driven by a large language model. Summary of the invention

[0008] The purpose of the present invention is to propose a large language model driven root cause analysis method for high carbon emission processes in the textile industry to solve the problems raised in the background technology. The present invention effectively reduces manual intervention and significantly improves the accuracy and reliability of root cause analysis of high carbon emission processes in sustainable textile production.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] A root cause analysis method for textile high carbon emission processes driven by a large language model, more specifically a root cause analysis method for textile high carbon emission processes in sustainable production based on a large language model (referred to as "large model"), the method comprising the following steps:

[0011] S1. Predefine production process data structure and semantic text representation method: predefine a new process data structure, the process data structure is "{case id, activity, timestamp, root cause}"; and propose a method for converting manufacturing process event logs into a predefined semantic text representation method, the predefined semantic text representation method is "start activity→end activity (result)"; for example, the transition from process 1 to process 2, with a frequency of 2 and a performance attribute of 150, can be represented as "process 1→process 2 (frequency=2, performance=150)";

[0012] A knowledge graph G is introduced, which is represented as a triple {H, R, T} storage. The root cause information corresponding to each process node is processed in a graphical manner, the causal knowledge related to the production process nodes is structured, and the task classification and domain knowledge storage are completed.

[0013] S2. Design a proximity retrieval method based on production process data: Utilize the large amount of historical process log data and specific expert knowledge graphs stored in the manufacturing management system to provide the possibility for precision auxiliary decision-making tasks, design a proximity retrieval method based on production process data, calculate the similarity of process data, and enhance the accuracy and credibility of auxiliary decision-making result generation and confidence estimation;

[0014] S3. Two-stage confidence assessment based on neighbor retrieval and confidence assessment: A two-stage confidence assessment scheme combining neighbor retrieval and confidence assessment is proposed to ensure a balance between the relevance of retrieved cases and the credibility of predicted root causes during neighbor retrieval, prompting the large language model to plan the problem in detail, decomposing complex problems of intelligent and automated mining of industrial processes into two manageable sub-problems, and obtaining the final decision-making assistance results.

[0015] Preferably, S2 specifically includes the following contents:

[0016] The data structure after the predefined large model processing is: D = {e i =(h i , a i ,s i , G i ); i = 1, 2, ..., n max}, including the historical process representation e i (represents the production event log data within a unit time range); each historical process contains a unique identifier index parameter h after processing i (Use fixed-dimensional numbers to represent data and make full use of context); the corresponding process semantic text representation a i (For example: "Process 1 → Process 2 (frequency = 2, performance = 150), ..."); Large model reasoning state information si (e.g.: normal or abnormal (high power consumption, material waste, etc.)); production resource triples G related to auxiliary decision making i (Other knowledge supplements are used for graph construction and carbon emission analysis queries);

[0017] A new similarity formula is designed to calculate the similarity of process data. The specific formula is:

[0018]

[0019] Among them, |t(a)-t(b)| represents the time correlation of process data; ||ab|| 2 Represents the Euclidean distance between data, and obtains the top-k process cases with the highest similarity of S(a, b) as the few-case strategy prompts for the large language model; a and b represent the textual representation of production logs in two unit time periods respectively; uses proximity retrieval to compare the similarity of data and obtains the top k cases related to the problem for subsequent reasoning support.

[0020] Preferably, S3 specifically includes the following contents:

[0021] A two-stage confidence evaluation scheme combining proximity retrieval and confidence evaluation is designed to achieve balanced evaluation value scores. The specific formula is expressed as:

[0022]

[0023] Where ω(i) represents the importance weight function of the sampled cases; E(p)) represents the mapping function between the retrieval case relevance and the prediction result reliability score;

[0024] For each query, multiple sampling analyses n are set, and the i-th sampled retrieval case is scored for proximity retrieval relevance. The specific formula is:

[0025]

[0026] Where n represents the number of sampling times; k j Indicates the number of similar historical cases retrieved and summarized by the large language model; is the evaluation score for the jth historical case sampled at the i-th time, indicating whether the current case has a supporting role for the current problem;

[0027] For the root cause analysis prediction evaluation of industrial process intelligent automation mining, multiple dimensional influencing factors are taken into account, including: information authenticity, information granularity and historical process support; five gradients are designed for scoring (scores range from 0-10), and the key nodes include non-correlation, correlation with more positive cases and fewer negative cases, and strong correlation. The specific formula is:

[0028]

[0029] Among them, E(p) is the core component of the two-stage confidence assessment scheme, which is used to calculate the mean of the root cause prediction score based on historical references; n i ′ is the i-th root cause analysis summarized by the large language model based on the retrieval results of each question instruction; k j ′ is the impact factor of the jth dimension; p′ j is the score of the i-th dimension influencing factor in the i-th root cause analysis.

[0030] Compared with the prior art, the present invention provides a root cause analysis method for textile high carbon emission processes driven by a large language model, which has the following beneficial effects:

[0031] The method for root cause analysis of high carbon emission processes in sustainable textile production based on a large language model proposed in the present invention realizes the structural predefinition, storage, and management of production process data and the root cause analysis of each high carbon emission process, ensuring the transparency and accurate traceability of carbon emissions at the process level. This method can be used to conduct traceability analysis of carbon emission problems and assist in process optimization, which helps to discover the carbon emission influencing factors of each specific process and thus improve the production process. The improved method proposed in the present invention helps to improve the green and sustainable production process of the industry and achieve the long-term goal of sustainable development. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is an overall flow chart of the root cause analysis method for textile high carbon emission processes driven by the language model mentioned in Example 1 of the present invention;

[0033] Figure 2 This is a schematic diagram of the two-stage confidence evaluation mentioned in Example 1 of the present invention;

[0034] Figure 3 This is a Directly-Follows Graph (DFG) model generated by the present invention using a process mining algorithm for log data. DETAILED DESCRIPTION

[0035] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are described in further detail below.

[0036] The present invention proposes a language model-driven root cause analysis method for textile high carbon emission processes. First, in order to meet the input requirements of the large model, the production process data structure and semantic text representation method are predefined to summarize and textualize the production process log data and graph data related to the domain knowledge stored in the local information management system. Secondly, the proximity retrieval method based on the production process data is used to match similar cases to form the support basis for the subsequent large model analysis. Finally, a two-stage confidence evaluation is performed based on the proximity retrieval case and the large model root cause analysis results, and the decision result is output.

[0037] Among them, the top priority of industrial process decision support is to use industrial process mining methods to analyze the influencing factors related to the abnormal state of the current manufacturing process data, such as the distribution of energy carbon emissions, process efficiency, rationality of tasks and personnel allocation. The decision support results obtained by traditional root cause mining relying on manpower cannot be effectively guaranteed in terms of efficiency and accuracy. The large model technology is affected by the difficulty of the manufacturing process task and outputs an indistinguishable illusion solution, which cannot play an accurate decision support effect. Therefore, this solution uses historical neighbor search to predict the root cause analysis, and the predicted results serve as the basis for subsequent large model evaluation and the source of confidence evaluation. The two-stage confidence evaluation is used to improve the accuracy and effectiveness of decision support. The goal of this step is to evaluate the model's ability to analyze and reason about the current problem instructions, and whether there is enough evidence to support its judgment. Finally, the confidence evaluation score of whether the information retrieved by the large model according to the problem instruction has the support to infer the root cause is output. If the retrieved information cannot support the subsequent inference and has a misleading effect, the root cause analysis results generated by the large model based on this type of retrieved knowledge will also be less credible. This evaluation score can help prevent the model from being overconfident in the absence of sufficient information, thereby avoiding misleading engineers.

[0038] See also Figure 1 The present invention proposes a root cause analysis method for high carbon emission processes in sustainable textile production based on a large language model, comprising the following steps:

[0039] (1) Predefined production process data structure and semantic text representation method

[0040] In industrial information management systems, processing large-scale, variable-length process event log data is a challenging task. Especially when these event logs interact with large models, data redundancy and low accuracy often occur. The traditional process log data format usually follows the agreed structure: {case id, activity, timestamp}. However, when log data and graph data interact with large models, this format often leads to unnecessary waste of context length, inefficient data processing, and invalid process data text representation. To solve this problem, this scheme predefines a new process data structure {case id, activity, timestamp, root cause} and proposes a method to convert manufacturing process event logs into a predefined semantic text representation. The predefined semantic text representation method is "start activity → end activity (result)". For example, the transition from process 1 to process 2 with a frequency of 2 and a performance attribute of 150 can be represented as "process 1 → process 2 (frequency = 2, performance = 150)".

[0041] In addition, traditional process mining techniques have limitations in providing decision support for the manufacturing industry, mainly relying on the log structure {case id, activity, timestamp}, and therefore cannot fully utilize other relevant production data. Specifically, the limitations of traditional process mining techniques include the lack of comprehensive consideration of key factors in the production process, such as resource consumption, carbon emissions, and other environmental factors. The reliance on a fixed log structure makes it difficult to adapt to the changing needs of different industries and production environments. It is difficult to handle large-scale, multivariate production data, resulting in low analysis efficiency. The lack of an effective data representation method limits direct interaction with advanced technologies such as large language models. To address this issue, this scheme introduces a knowledge graph G, represented as a triple {H, R, T} storage, to graphically process the root cause information corresponding to each process node. This processing structures the causal knowledge associated with the production process nodes, facilitating task classification and storage of domain knowledge.

[0042] (2) Proximity retrieval method based on production process data

[0043] See also Figure 2 ,The important obstacle to root cause analysis tasks in the field of manufacturing process mining lies in the specific manufacturing background knowledge. The starting point of the design of large model technology is general field application, and the technology lacks the domain knowledge necessary for a specific manufacturing field. However, the large amount of historical process log data and specific expert knowledge graphs stored in the manufacturing management system provide the possibility for precision auxiliary decision-making tasks, such as root cause analysis. Therefore, this scheme enhances the accuracy and credibility of auxiliary decision-making result generation and confidence estimation by designing the corresponding neighbor retrieval method.

[0044] The data structure after the predefined large model processing is D = {e i =(h i , a i ,s i , G i ); i = 1, 2, ..., n max}, including the historical process representation e i (represents the production event log data within a unit time range), each historical process contains a unique identifier index parameter h after processing i (using fixed-dimensional numbers to represent data and making full use of context), the corresponding process semantic text representation a i (e.g.: "Process 1 → Process 2 (frequency = 2, performance = 150), ..."), large model inference state information s i (e.g. normal or abnormal (high power consumption, material waste, etc.)), assisting decision-making related production resource triples G i (Other knowledge supplements are used for graph construction and carbon emission analysis query.) Considering that similar data pairs are affected by the Euclidean distance factor, the time similarity between data pairs should also be considered. For this purpose, this solution designs a new similarity formula to calculate the similarity of process data.

[0045]

[0046] Among them, |t(a)-t(b)| represents the time correlation of process data, ||ab|| 2 Represents the Euclidean distance between data, and obtains the top-k process cases with the highest similarity of S(a,b) as partial case strategy prompts for the large model. This retrieval method combines similar cases of domain knowledge to generate accurate auxiliary decision results that integrate content similarity and time similarity, where a and b represent the textual representation of production logs in two unit time periods. Use proximity retrieval to compare the similarity of data and obtain the top k cases related to the problem for subsequent reasoning support.

[0047] (3) Two-stage confidence evaluation based on proximity retrieval

[0048] One of the obstacles to the generation of accurate decision methods by large language models is the existence of well-disguised hallucinations. Accurate auxiliary decision-making requires a balance between the relevance of the retrieved cases and the credibility of the predicted root causes during the neighboring retrieval process. Therefore, this scheme proposes a two-stage confidence assessment scheme that combines neighboring retrieval and confidence assessment.

[0049] The method shown is to encourage the model to plan the problem carefully, decompose the complex problem of intelligent automation mining of industrial processes into two manageable sub-problems, and obtain the final auxiliary decision-making results. The two-stage confidence evaluation achieves the balance of the evaluation value scores, as shown below:

[0050]

[0051] Among them, ω(i) represents the importance weight function of the sampled case, Represents the mapping function between the retrieval case relevance and the prediction result reliability score. Specifically, for each query, multiple sampling analyses are set n, and for the i-th sampled retrieval case k j Perform proximity search relevance scoring The specific formula is as follows:

[0052]

[0053] Among them, this problem can be transformed into a binary classification problem, using the reasoning ability of the large model to perform a binary classification score (Yes / No) for each case, mapped to {0,1}; n represents the number of sampling times; k j Indicates the number of similar historical cases retrieved and summarized by the large model; It represents the evaluation score of the jth historical case sampled at the i-th time, indicating whether the current case has a supporting role for the current problem;

[0054] Similarly, for the root cause analysis and prediction evaluation of intelligent automation mining of industrial processes, multiple dimensional influencing factors need to be considered, including: information authenticity, information granularity, and the degree of historical process support. Therefore, this solution designs five gradient scoring methods (scores range from 0-10), and the key nodes include non-relevant, relevant and retrieved positive cases with fewer negative cases, and strong correlation, as shown below:

[0055]

[0056] Among them, E(p) is the core component of the two-stage confidence evaluation scheme, which reflects the purpose of evaluating whether the root cause predicted by the model is reasonable and its degree of relevance to historical events. Formula (4) evaluates the reliability of the predicted root cause analysis results by considering influencing factors in multiple dimensions, such as the authenticity of information, the granularity of information, and the support level of historical processes (5 gradient scores for each dimension from 0 to 10, with key points including irrelevance, relevance of more positive cases than negative cases in the retrieval, and strong relevance). Formula (4) is used to calculate the mean of the root cause prediction scores based on historical references. Specifically, it is the mean of the multidimensional influencing factor scores generated by each root cause analysis obtained through retrieval reasoning according to the question instructions. Among them, n i′ is the i-th root cause analysis summarized by the large model based on the retrieval results of each problem instruction; k j ′ is the impact factor of the jth dimension, p′ j is the score of the i-th dimension influencing factor in the i-th root cause analysis.

[0057] Based on the above content, the language model-driven root cause analysis method for high carbon emission processes in the textile industry proposed in the present invention is described below in conjunction with specific examples. The specific content is as follows.

[0058] Embodiment 1:

[0059] A root cause analysis method for textile high carbon emission processes driven by a large language model includes the following steps:

[0060] (1) Predefined production process data structure and semantic text representation method

[0061] First, the process event log files stored in the information system are processed by the method shown in this scheme and stored locally to form a long-term memory library of the large model for historical case retrieval and as input for prompt engineering corpus. The partial textile production carbon emission data structure used in this scheme is shown in Table 1.

[0062] Table 1 Part of the event log data structure

[0063]

[0064] As shown in Table 1, Table 1 describes the data related to carbon emissions in the textile production process. Processing and storing this type of data is a key source for improving the reliability of large model reasoning. The direct follow-up graph (DFG) model is generated using the process mining algorithm for log data. Figure 3 As shown in the figure, nodes represent processes, edges represent direct follow-up relationships between processes, and weights can be switched to represent the frequency, duration, and total carbon emissions of processes. The process representation method proposed in this scheme converts the above DFG model into process semantic text, and the converted semantic text represents: "Warehousing → Weaving (frequency = 79, performance = 1299.494); Weaving → Printing and Dyeing (frequency = 38, performance = 2490.000); ...; Activity head: curling; root cause analysis: normal, benefit (total carbon emissions: 5362.262)". Other knowledge is used to construct a knowledge graph, which is used to integrate the equipment and energy consumption data knowledge and management related to the state as a basis for supporting the judgment of auxiliary analysis results.

[0065] (2) Proximity retrieval method based on production process data and two-stage confidence evaluation

[0066] By using the knowledge graph to search for historical cases, we can obtain carbon emission evidence of various types of textile production with abnormally high carbon emission processes, for example: "The search case number is 1002, the textile type is 80% wool + 12% nylon + 8% spandex blended jersey, the dyeing process in the processing flow is abnormal (high power consumption), the total carbon emission of electricity is 1634.264KgCO2e, the equipment model is BRAZZO, the equipment type is flow-equalizing dyeing machine, the equipment operation carbon emission is 1535.055KgCO2e, the air conditioning equipment ... Carbon emissions from lighting equipment are 23.271KgCO2e, carbon emissions from auxiliary equipment are 24.494KgCO2e, carbon emissions from steam are 2,015.070KgCO2e, carbon emissions from dyes are 105KgCO2e, consumption of auxiliaries is 802.92KgCO2e and carbon emissions from water energy are 71.51KgCO2e; Retrieval case 2, order number 1040, textile type is horizontal striped sweat cloth, and the dyeing process in the processing flow presents an abnormal state (waste of auxiliaries), etc." Then, the large model performs reasoning based on the previous k historical cases and the problem instructions in the prompt word template, summarizes the root cause analysis results, and further generates the reasoning confidence score E(p) to characterize the credibility of the model reasoning results.

[0067] Finally, based on the above two scores, a two-stage confidence assessment method will be used to generate the final confidence score for reference in reasoning, which shows that the reasoning results have reliable case support and the results themselves are highly credible. Finally, the data will be sorted and returned according to the preset output template for decision support. For example, the final output results of this case are summarized as follows, "Problem: Root cause analysis of high carbon emission processes in key processes of abnormal carbon emissions in textile production; Results: 1-Dyeing process-auxiliary agent consumption (auxiliary agent: lubricant, carbon emission factor is 2.95KgCO2e / kg)-Confidence assessment: 9.3; 2-Dyeing process-high power consumption (equipment: BRAZZO; average power consumption: 1,895.130kw·h)-Confidence assessment: 8.4..."

[0068] It should be noted that, in the present invention patent, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0069] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A root cause analysis method for textile high carbon emission processes driven by a large language model, characterized in that: The following steps are involved: S1. Predefine production process data structure and semantic text representation method: predefine a new process data structure, the process data structure is "{case id, activity, timestamp, root cause}"; and propose a method for converting manufacturing process event logs into a predefined semantic text representation method, the predefined semantic text representation method is "start activity → end activity (result)"; introduce a knowledge graph G, represented as a triple {H, R, T} storage, process the root cause information corresponding to each process node in a graphical manner, structure the causal knowledge related to the production process nodes, and complete the task classification and storage of domain knowledge; S2. Design a proximity retrieval method based on production process data: Utilize the large amount of historical process log data and specific expert knowledge graphs stored in the manufacturing management system to provide the possibility for precision auxiliary decision-making tasks, design a proximity retrieval method based on production process data, calculate the similarity of process data, and enhance the accuracy and credibility of auxiliary decision-making result generation and confidence estimation; S3. Two-stage confidence assessment based on neighbor retrieval and confidence assessment: A two-stage confidence assessment scheme combining neighbor retrieval and confidence assessment is proposed to ensure a balance between the relevance of retrieved cases and the credibility of predicted root causes during neighbor retrieval, prompting the large language model to plan the problem in detail, decomposing complex problems of intelligent and automated mining of industrial processes into two manageable sub-problems, and obtaining the final decision-making assistance results.

2. According to claim 1, a large language model driven root cause analysis method for textile high carbon emission processes is characterized in that: The S2 specifically includes the following contents: The data structure after processing by the predefined large language model is recorded as: D={e i =(h i ,a i ,s i ,G i )} Where i = 1, 2, ..., n max ;e i It represents the historical process; h i Is a unique identifier index parameter; a i It is the textual representation of the process corresponding to the historical process; i Reasoning state information for large language models; G i To assist decision-making related production resource triples; A new similarity formula is designed to calculate the similarity of process data. The specific formula is: Among them, |t(a)-t(b)| represents the time correlation of process data; ||ab||2 represents the Euclidean distance between data, and the top-k process cases with the highest similarity to S(a, b) are obtained as hints for the few-case strategy of the large language model; a and b represent the textual representation of production logs in two unit time periods respectively; proximity retrieval is used to perform similarity comparison on the data to obtain the top k cases related to the problem for subsequent reasoning support.

3. According to claim 2, a large language model driven root cause analysis method for textile high carbon emission processes is characterized in that: The S3 specifically includes the following contents: A two-stage confidence evaluation scheme combining proximity retrieval and confidence evaluation is designed to achieve balanced evaluation value scores. The specific formula is expressed as: Where ω(i) represents the importance weight function of the sampled cases; E(p)) represents the mapping function between the retrieval case relevance and the prediction result reliability score; For each query, multiple sampling analyses n are set, and the i-th sampled retrieval case is scored for proximity retrieval relevance. The specific formula is: Where n represents the number of sampling times; k j Indicates the number of similar historical cases retrieved and summarized by the large language model; is the evaluation score for the jth historical case sampled at the i-th time, indicating whether the current case has a supporting role for the current problem; For the root cause analysis and prediction evaluation of intelligent automation mining of industrial processes, multiple dimensional influencing factors are taken into account, including: information authenticity, information granularity, and the degree of historical process support; five gradients are designed for scoring, and the key nodes include non-correlation, correlation with more positive cases and fewer negative cases, and strong correlation. The specific formula is: Among them, E(p) is the core component of the two-stage confidence assessment scheme, which is used to calculate the mean of the root cause prediction score based on historical references; n i ′ is the i-th root cause analysis summarized by the large language model based on the retrieval results of each question instruction; k j ′ is the impact factor of the jth dimension; p′ j is the score of the i-th dimension influencing factor in the i-th root cause analysis.

Images