Manufacturing process intelligent decision-making-oriented fine-tuning large language model integration system

Abstract

The invention discloses a manufacturing industry intelligent decision-making-oriented fine-tuning large language model integration system, and the system achieves the capture of different feature subspace decision-making information through a plurality of parallel manufacturing industry fine-tuning large language models, and generates a plurality of preliminary decision-making schemes. Then modeling is directly carried out on the probability of the optimal candidate preliminary decision scheme to obtain a candidate sequence, and the candidate sequence is realized under the guidance of multiple metric heads; the top-ranked candidate schemes are then fused using the best fine tuning LLM to reduce deviation information in the final decision scheme. In conclusion, the fine-tuning large language model integration system for intelligent decision making in the manufacturing process aims at improving the context analysis capability of a manufacturing decision making system on complex association of the manufacturing system, and the probability of decision making deviation is reduced by integrating the advantages of all aspects through the sorting strategy and the fusion strategy.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent decision-making, and specifically relates to an integrated system of fine-tuned large language models for intelligent decision-making in the manufacturing process. Background Art

[0002] Decision-making, as an indispensable part of manufacturing systems, plays a crucial role in aspects such as production, operations management, and continuous improvement. In the manufacturing field, decision-makers usually need to rely on their experience and expertise to derive multiple alternative solutions based on the specific on-site situation and make decision selections on this basis. These decision selections run through the entire production and manufacturing cycle, from the design and scheduling of production lines, product processing and manufacturing to equipment operation and maintenance. Efficient and accurate decision-making can greatly affect the entire production process, helping manufacturing enterprises simplify the decision-making process, reduce uncertainties, improve efficiency, and enhance overall performance. Especially when unforeseen and sudden quality defects, equipment failures, or other unexpected events occur during the production process, quick and accurate decision-making can effectively avoid the decline of product quality, production stagnation, and even safety accidents. However, it is almost impossible to achieve completely continuous optimal decision-making, and the manufacturing decision-making process will inevitably be affected by various internal and external factors, resulting in decision-making biases. Therefore, constructing a manufacturing intelligent decision-making framework that can effectively alleviate manufacturing decision-making biases is crucial for enterprises to cope with complex production challenges, ensure stable product quality, and maintain high production capacity.

[0003] The reasons for decision-making biases in the manufacturing field can be divided into two parts: objective and subjective, such as Figure 1As shown. Objectively speaking, the complexity of modern manufacturing systems is continuously increasing, including but not limited to the complex correlations in aspects such as equipment, information flow, process flows, and human-machine interactions, which poses new requirements for manufacturing decision-making methods. With the booming development of artificial intelligence technology, advanced artificial intelligence technologies such as machine learning and deep learning are gradually integrated into manufacturing decision support systems. These technologies are different from traditional rule-based decision trees and expert systems, which rely on pre-set rule sets or templates, while artificial intelligence-driven methods automatically identify patterns and make predictions through training models, thereby achieving an understanding of manufacturing data. The application of artificial intelligence in manufacturing decision-making is mainly reflected in two paradigms: data-driven and knowledge-driven. Data-driven methods focus on mining specific patterns from large-scale data and are applicable to scenarios such as condition monitoring and predictive maintenance. Although these methods can assist in decision-making, they usually do not directly generate decision suggestions; knowledge-driven methods utilize knowledge graphs or pre-trained models to derive decision-making schemes through in-depth analysis of manufacturing knowledge, thus making up for the deficiencies of data-driven methods. The rise of large language models has significantly improved the depth of understanding of manufacturing domain expertise. With its excellent semantic parsing ability, it can effectively identify and extract implicit semantic information in the manufacturing process. By integrating large language models into the manufacturing decision-making framework, the intelligence level of the manufacturing decision-making framework can be further enhanced.

[0004] Subjectively, as the complexity of manufacturing systems continues to increase, the frequency and types of potential problems also expand accordingly. This complexity increases the number of uncertain factors in the manufacturing process. Although existing manufacturing decision-making frameworks can assist engineers in making decisions, most of these frameworks are built on a closed-domain knowledge system, and their construction and optimization highly rely on the experience and expertise of domain experts. Due to the different knowledge backgrounds, experiences, and preferences of different experts, their understandings and decision-making schemes for the same manufacturing problem may vary, which directly affects the consistency and reliability of decision-making. Especially in a highly interconnected complex manufacturing environment involving equipment operation and process flows, this subjective difference will gradually accumulate in the manufacturing decision-making system, thus affecting the final product quality and production stability. In addition, when engineers interact with manufacturing decision-making frameworks, they face huge information processing pressure, and limited cognitive abilities and time resources make it difficult for them to make optimal decisions in the case of information overload, which further exacerbates the risk of decision-making bias. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a fine-tuned large language model integration system for intelligent decision-making in the manufacturing process, aiming to improve the context parsing ability of the manufacturing decision-making system for the complex correlations of manufacturing systems, and through sorting strategies and fusion strategies, integrate the advantages of various aspects, thereby reducing the probability of decision-making bias.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An integrated system of fine-tuned large language models for intelligent decision-making in the manufacturing process, including a fine-tuned large language model LLM, a DWMOE module, and an AFDM module;

[0008] A plurality of the fine-tuned large language models LLM are provided, and each fine-tuned large language model LLM i generates a preliminary decision-making plan PD according to a given decision-making requirement X i ;

[0009] The DWMOE module realizes the final sorting of all preliminary decision-making plans by defining a comprehensive sorting function Rank(PD i ), and performing weighted sorting based on a plurality of sorting metric functions {f 1 (i), f 2 (i), …, f m (i)}; m is the number of sorting metric functions;

[0010] The AFDM module selects the top K preliminary decision-making plans as candidate decision-making plans through an adaptive fusion decision-making strategy and splices them to form a fusion instruction;

[0011] Taking the fine-tuned large language model LLM corresponding to the preliminary decision-making plan ranked first as a fusion device, a final manufacturing decision-making plan is generated based on the fusion instruction.

[0012] Furthermore, the fine-tuned large language model LLM is obtained by using the LORA fine-tuning method based on a general-domain large language model; the LORA fine-tuning method is as follows:

[0013] Freeze the weight matrix W of the general-domain large language model, and introduce a trainable low-rank matrix ΔW; the low-rank matrix ΔW is represented by the product of two small matrices r << d and r << k:

[0014] ΔW = AB T

[0015] where: r(A) = R(B) = r;

[0016] Given a manufacturing decision-making requirement Q, represent the manufacturing decision-making requirement Q as an input feature vector Map the feature vector x Q through a linear transformation of a certain layer for feature mapping, and the fine-tuned weight matrix W' is represented as:

[0017] W' = W + ΔW = W + AB T

[0018] The linear transformation result h of the manufacturing decision requirement Q Q is as follows:

[0019] h Q = W'x Q = (W + AB T )x Q = Wx Q + A(B T x Q )

[0020] where: A(B T x Q ) is the feature adjustment of the manufacturing decision requirement Q

[0021] Furthermore, multiple different fine-tuned large language models {LLM 1 , LLM 2 , …, LLM n} are obtained by using the LORA fine-tuning method based on different general domain large language models respectively; each fine-tuned large language model LLM i combines its own pre-trained general knowledge and manufacturing knowledge to deeply analyze the manufacturing decision requirement Q, realizes multi-dimensional semantic capture of manufacturing knowledge, and deeply understands the complex associations of the manufacturing system; at the same time, the parallel fine-tuned large language model LLM performs multi-dimensional feature transformation and extraction in multiple feature subspaces, obtains relevant manufacturing decision information in different feature subspaces, and generates preliminary decision-making schemes {PD 1 , PD 2 , …, PD n} with mutually related but different decision-making information; n is the number of fine-tuned large language models

[0022] Furthermore, the DWMOE module includes a semantic analysis part and a mixture of experts part. The semantic analysis part is used to realize semantic feature extraction, and the mixture of experts part is used to realize the construction of a probability distribution model and realize candidate ranking under the guidance of multiple metric heads

[0023] Furthermore, the semantic analysis part includes a semantic parser S and a non-linear feature processor M; the semantic parser S uses a pre-trained model that can effectively improve the quality of semantic features; the method for the semantic analysis part to realize semantic feature extraction is: splicing the decision requirement Q with the preliminary decision-making scheme {PD 1 , PD 2 , …, PD n} to obtain the input representation x input = [Q; PD i , and using the input representation x inputInput into the semantic parser S; extract the feature representation h of the last layer of the pre-trained model RoBERTa (x input ), and input it to the non-linear feature processor M for non-linear feature processing to obtain the non-linear feature x hidden = MLP(h RoBERTa (x input ))

[0024] Furthermore, the mixture-of-experts part consists of a sparse mixture of experts and multiple metric heads; the multiple metric heads include at least two metric heads, and each metric head uses a different evaluation mechanism to evaluate the same text to achieve feature complementarity;

[0025] For each metric head H, introduce a set of sparse mixture-of-experts models E H ={e H,1 , e H,2 , …, e H,n}, and each set of models consists of n experts; for each metric head, calculate the attention weights based on the intermediate feature representation x hidden , and each expert e H ={e H,1 , e H,2 , …, e H,n} in the expert set E H,i will get a weight α H,i , expressed as:

[0026]

[0027] where: score is the evaluation function in the attention mechanism, used to measure the correlation between x hidden and each expert e H,i , and the dot product is used for evaluation scoring:

[0028]

[0029] where: is the weight matrix;

[0030] Perform weighted summation on the outputs of the experts to obtain the score s H of the metric head H:

[0031]

[0032] Construct a loss function based on the scores of all metric heads:

[0033]

[0034] where: is the loss function; λ i is the weight of the metric head Hi; is the metric loss of the metric head Hi; s i is the score of the metric head Hi; y is the label score; p is the number of metric heads;

[0035] Based on the weights of the metric heads, the scores of all metric heads are weighted and measured:

[0036]

[0037] where: s is the metric score of each preliminary decision scheme.

[0038] The beneficial effects of the present invention are as follows:

[0039] Intelligent decision-making is the core element to activate the value of industrial knowledge and promote the full-scenario empowerment of the manufacturing industry. However, traditional manufacturing decision-making methods cannot fully clarify the complex relationships among the various elements of the manufacturing industry, and there are often problems with decision-making biases. As a new generation of productivity tool, the large language model has powerful context semantic parsing capabilities. Therefore, the present invention proposes a fine-tuned large language model integration system for intelligent decision-making in the manufacturing industry. This system captures decision-making information in different feature subspaces through multiple parallel large language models fine-tuned for the manufacturing industry and generates multiple preliminary decision-making schemes; then directly models the probability of the best candidate preliminary decision-making scheme to obtain a candidate ranking, and realizes the candidate ranking under the guidance of multiple metric heads; immediately afterwards, the best fine-tuned LLM is used to fuse the candidate schemes ranked among the top to reduce the deviation information in the final decision-making scheme. In summary, the fine-tuned large language model integration system for intelligent decision-making in the manufacturing process of the present invention aims to improve the context parsing ability of the manufacturing decision-making system for the complex relationships of the manufacturing system, and through the ranking strategy and the fusion strategy, integrate the advantages of all aspects, thereby reducing the probability of the occurrence of decision-making biases.

[0040] To solve the problem of manufacturing decision-making bias, the present invention proposes a fine-tuned large language model integration system (LLM-MANUF) for intelligent decision-making in the manufacturing industry. Specifically, the LLM-MANUF system includes multiple large language models with small parameter numbers and fine-tuned with a manufacturing corpus, aiming to improve the context parsing ability of the manufacturing decision-making system for the complex relationships of the manufacturing system, and through the ranking strategy and the fusion strategy, integrate the advantages of all aspects, thereby reducing the probability of the occurrence of decision-making biases. The main technical content and technical effects of the present invention can be summarized as follows:

[0041] (1) The present invention designs an LLM integration system LLM-MANUF. By using the LLM-MANUF system to alleviate the problem of manufacturing decision-making bias, multiple fine-tuned LLM models are used to jointly analyze the same decision-making task, and then the ranking and fusion strategies are applied to integrate the advantages of these models, thereby generating an optimal decision-making scheme.

[0042] (2) The present invention studies a multi-LLM fine-tuning model for intelligent decision-making in manufacturing. A manufacturing corpus is obtained by sorting out manufacturing-related documents. Several LLMs with low computing resource requirements are selected, and the LoRA fine-tuning technology is used to specifically fine-tune these large language models based on the aforementioned manufacturing-specific corpus, obtaining a multi-LLM manufacturing fine-tuning model.

[0043] (3) The present invention introduces a Dynamic Weighted Mixture of Experts Ranking Method (DWMOE). This ranking method obtains candidate rankings by modeling the probabilities of the best candidates. Semantic feature extraction is achieved using a semantic parser, a probability distribution model is constructed based on a dynamic weighted mixture of experts system, and candidate rankings are realized under the guidance of multiple metric heads.

[0044] (4) The present invention proposes an Adaptive Fusion Decision Strategy (AFDS). In this fusion decision strategy, the best LLM can be dynamically selected as the fusion device according to the specific requirements of the manufacturing decision-making scenario to effectively integrate the best candidate solutions. Through this process, not only can the advantages of each candidate solution be synthesized, but also the accuracy and reliability of the final decision-making solution can be ensured, thus significantly reducing the deviation in manufacturing decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] To make the objectives, technical solutions, and beneficial effects of the present invention clearer, the following drawings are provided for illustration:

[0046] Figure 1 Problems existing in the existing manufacturing decision-making methods;

[0047] Figure 2 Framework diagram of the fine-tuned large language model integration system for intelligent decision-making in manufacturing according to the present invention;

[0048] Figure 3 Schematic diagram of the generation of the preliminary decision-making plan;

[0049] Figure 4 Schematic diagram of the LoRA fine-tuning method;

[0050] Figure 5 Schematic diagram of the dynamic weighted mixture of experts ranking method DWMOE;

[0051] Figure 6 Flowchart of data preprocessing;

[0052] Figure 7 Bar chart of the decision-making time consumption of different models. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments cited are not intended to limit the present invention.

[0054] 1. Related work

[0055] 1.1. Decision-making research in the manufacturing field

[0056] Manufacturing decision-making bias refers to the situation in a manufacturing environment where, due to the cognitive limitations of decision-makers or decision-making systems, it is difficult to comprehensively understand the complex interrelationships within the manufacturing system, resulting in deficiencies in information acquisition, processing, and analysis. Current research focuses on alleviating decision-making bias by enhancing the decision-making ability of manufacturing decision-making methods or standardizing the manufacturing decision-making processes of decision-makers.

[0057] In terms of enhancing the ability of manufacturing decision-making methods, research mainly focuses on strengthening the understanding of the complex associations in the manufacturing system, which can be achieved through two approaches: data-driven and knowledge-driven. Among them, the data-driven approach only focuses on monitoring and diagnosing the state of production and manufacturing and cannot directly assist users in decision-making; knowledge-driven manufacturing decision-making emphasizes more on guiding the decision-making process based on accumulated professional knowledge and experience. This method not only aims to provide direct decision-making suggestions to users but also promotes in-depth understanding and optimization of the manufacturing system through systematic management and induction of professional knowledge in the manufacturing field. Although deep learning models such as graph neural networks and BERT have demonstrated remarkable semantic parsing capabilities in the field of natural language processing, when applied to manufacturing systems, their understanding of manufacturing-specific semantics is still relatively shallow. In contrast, as a model of the new generation of artificial intelligence, large language models provide a new approach to solving the above problems through their powerful context understanding and reasoning capabilities. Large language models provide a feasible and effective solution for parsing the complex associations in manufacturing systems through their excellent context understanding and reasoning capabilities, as well as learning of deep semantics in the manufacturing field.

[0058] In the prior art, there are also studies focusing on standardizing the decision-making processes of decision-makers, thereby reducing the possibility of decision-making bias by decision-makers. Multiple studies have shown that by integrating multiple decision-making models to form a comprehensive decision-making system, the reliability and accuracy of decision-making can be significantly improved.

[0059] Although existing studies have achieved certain results in alleviating the problem of manufacturing decision bias, they all have some limitations. Therefore, this embodiment proposes a method based on fine-tuning a large language model, aiming to achieve multi-dimensional semantic capture of manufacturing knowledge, providing strong support for in-depth understanding of the complex associations of manufacturing systems. On this basis, this embodiment further constructs a manufacturing intelligent decision integration system that integrates multiple fine-tuned LLMs to integrate decision information from different feature subspaces. The system can comprehensively analyze the key factors in the manufacturing process from multiple perspectives, thereby providing more accurate and reliable manufacturing decisions.

[0060] 1.2 Research on Ensemble Methods

[0061] A typical integrated system usually consists of three main parts: base model, ranking method and fusion method. The base model is responsible for generating preliminary candidate solutions; the ranking method ranks these candidate solutions according to specific requirements; the fusion strategy aims to integrate the ranked information to ensure the comprehensiveness and accuracy of the final output.

[0062] The main purpose of sorting is to arrange candidate objects in a certain order according to specific requirements. The performance of common sorting models is usually measured by loss functions. However, directly training the loss function between predicted values ​​and true values ​​may lead to a decrease in the generalization ability of the sorting model, especially when dealing with complex associations in the manufacturing process. In contrast, direct modeling of the probability of the best candidate can better maintain the generalization of the model. The probabilistic modeling method avoids direct reliance on ranking loss by estimating the probability of each candidate object becoming the optimal solution, thereby improving the adaptability of the model in different scenarios.

[0063] After the sorting method, an ordered list of candidate solutions is usually obtained. In order to ensure the completeness of the information, it is usually necessary to fuse the top-ranked solutions. The fusion strategy aims to integrate the advantages of multiple candidate solutions, avoid information omissions, and provide more comprehensive decision support. However, existing methods usually introduce new models when selecting a fuser, which may cause the fuser to relearn the context semantics. In contrast, selecting the base model that performs best in the task as the fuser has greater advantages. The base model has been fully trained in a specific task, has good semantic understanding capabilities, and can be directly applied to fusion tasks. In addition, it is also crucial to incorporate decision requirements into the input of the fuser.

[0064] 2. Fine-tuned large language model integration system for intelligent decision-making in manufacturing processes

[0065] like Figure 2As shown in the figure, the fine-tuned large language model integration system for intelligent decision-making in the manufacturing process of this embodiment includes a fine-tuned large language model LLM, a DWMOE module, and an AFDM module. Multiple fine-tuned large language models LLM are set, and each fine-tuned large language model LLM i generates a preliminary decision-making plan PD according to the given decision-making requirement X i . The DWMOE module realizes the final sorting of all preliminary decision-making plans by defining a comprehensive sorting function Rank(PD i ), and performing weighted sorting based on multiple sorting metric functions {f 1 (i), f 2 (i), …, f m (i)}; m is the number of sorting metric functions. The AFDM module selects the top K preliminary decision-making plans as candidate decision-making plans through an adaptive fusion decision-making strategy and splices them to form a fusion instruction. Using the fine-tuned large language model LLM corresponding to the preliminary decision-making plan ranked first as a fusion device, a final manufacturing decision-making plan is generated based on the fusion instruction.

[0066] 2.1. Generation of preliminary decision-making plan

[0067] The main role of the multi-fine-tuned LLM is to explore different feature spaces according to the decision-making requirement and generate a preliminary decision-making plan. In this process, it is mainly the fine-tuning process of the multi-LLM for intelligent decision-making in the manufacturing industry. In addition, it also includes the selection of the multi-LLM and the generation of the plan. Specifically as Figure 3 shown.

[0068] 2.1.1. LORA fine-tuning method

[0069] Large language models are usually pre-trained on general domain corpora, which makes them perform well in multi-task processing such as question answering and reasoning. However, due to the extensiveness and non-specificity of the pre-trained corpus, the model has insufficient knowledge reserves in the manufacturing field. To fill this knowledge gap, it is necessary to fine-tune the large language model for the manufacturing field. This fine-tuning process helps the model absorb and apply the professional knowledge in the manufacturing field to meet the needs of manufacturing users and improve the efficiency and accuracy of the large language model in manufacturing industry application scenarios.

[0070] This embodiment obtains the fine-tuned large language model LLM by using the LORA fine-tuning method based on the general domain large language model. LORA is an efficient lightweight fine-tuning technology, and this method mainly models the adjustment of weights by introducing a low-rank matrix. Specifically, as Figure 4 shown, the LORA fine-tuning method is as follows:

[0071] Freeze the weight matrix W of the general-domain large language model and introduce a trainable low-rank matrix ΔW; the low-rank matrix ΔW is represented by the product of two small matrices r << d and r << k:

[0072] ΔW = AB T

[0073] where: r(A) = r(B) = r;

[0074] Given a manufacturing decision requirement Q, represent the manufacturing decision requirement Q as an input feature vector The feature vector x Q Undergo feature mapping through a linear transformation of a certain layer, and the fine-tuned weight matrix W′ is expressed as:

[0075] W′ = W + ΔW = W + AB T

[0076] The linear transformation result h of the manufacturing decision requirement Q Q is:

[0077] h Q = W′x Q = (W + AB T )x Q = Wx Q + A(B T x Q )

[0078] where: A(B T x Q ) is the feature adjustment of the manufacturing decision requirement Q. The pre-trained knowledge of the model and new manufacturing low-rank information are included in this matrix.

[0079] 2.1.2. Preliminary Decision Model

[0080] Based on different general-domain large language models, respectively adopt the LORA fine-tuning method to obtain multiple different fine-tuned large language models {LLM 1 , LLM 2 , …, LLM n}; each fine-tuned large language model LLM i Combined with its own pre-trained general knowledge and manufacturing knowledge, deeply analyze the manufacturing decision requirement Q, realize multi-dimensional semantic capture of manufacturing knowledge, and deeply understand the complex associations of the manufacturing system; at the same time, the parallel fine-tuned large language model LLM performs multi-dimensional feature transformation and extraction in multiple feature subspaces, obtains relevant manufacturing decision information in different feature subspaces, and generates preliminary decision-making schemes {PD 1 , PD 2,…,PD n}; n is the number of fine-tuned large language models.

[0081] In this embodiment, five large language models that perform excellently in terms of performance, reproducibility, scalability, and popularity are carefully selected as the preliminary decision-making models. It should be noted that due to the limitations of computing power resources in the manufacturing industry, the large language models in this integrated system all select models with relatively small numbers of parameters, and the parameter scales of all models are kept consistent, about 7B. In addition, the number and parameter scale of the models can be flexibly adjusted according to the actual computing power resource conditions to adapt to the integration requirements of more models and diverse application scenarios, such as 2B, 13B, or 70B. The selected models include Llama3, qwen2.5, Gemma2, chatglm3, and Baichuan2. For large language models with multiple variants, the preferred version is the instruct version after instruction tuning, the second preferred version is the chat version, and finally the base model is selected. In terms of language selection, if a large language model is trained with a corpus such as English, the one fine-tuned with a Chinese corpus is preferred. The main features of the fine-tuned LLM are shown in Table 1.

[0082] Table 1 Details of the fine-tuned LLM

[0083]

[0084] Multiple fine-tuned LLMs are integrated in the LLM-MANUF system. Each LLM will combine its own pre-trained general knowledge and manufacturing knowledge to deeply analyze the manufacturing decision-making requirement Q, realize multi-dimensional semantic capture of manufacturing knowledge, and deeply understand the complex associations of the manufacturing system. At the same time, the parallel fine-tuned LLMs perform multi-dimensional feature transformation and extraction in multiple feature subspaces, obtain relevant manufacturing decision-making information in different feature subspaces, and generate preliminary decision-making schemes {PD 1 ,PD 2 ,…,PD n} that are mutually related but have different decision-making information.

[0085] 2.2. Dynamic Weighted Mixture of Experts Ranking Method (DWMOE)

[0086] The main goal of DWMOE is to rank the preliminary decision-making schemes {PD 1 ,PD 2 ,…,PD n}, perform sorting. The dynamic weighted mixture-of-experts sorting method obtains the candidate sorting based on modeling the probability of the best candidate. The semantic feature extraction is implemented using the RoBERTa semantic parser, the probability distribution model is constructed based on the dynamic weighted mixture-of-experts system, and the candidate sorting is achieved under the guidance of multiple metric heads. Specifically, as Figure 5 shown. That is, in this embodiment, the DWMOE module includes a semantic parsing part and a mixture-of-experts part. The semantic parsing part is used to implement semantic feature extraction, and the mixture-of-experts part is used to construct the probability distribution model, and the candidate sorting is achieved under the guidance of multiple metric heads.

[0087] (1) Semantic parsing part

[0088] The semantic parsing part includes a semantic parser S and a non-linear feature processor M. The semantic parser S uses a pre-trained model that can effectively improve the quality of semantic features. The pre-trained models include BERT, RoBERTa, and Deberta, etc., which have been proven to effectively improve the quality of semantic features. In this embodiment, RoBERTa is selected as the semantic parser S.

[0089] The method for the semantic parsing part to implement semantic feature extraction is as follows: In order to enable the semantic parser S to obtain better semantic features and make full use of the context information, the decision-making requirement Q is concatenated with the preliminary decision-making solutions {PD 1 , PD 2 , …, PD n} to obtain the input representation x input = [Q; PD i , and its specific input form is: [CLS]Query[SEP]PreliminaryDecision-Making. The input representation x input is input into the semantic parser S. The last layer feature representation h RoBERTa (x input ) of the pre-trained model is extracted and input to the non-linear feature processor M for non-linear feature processing to obtain the non-linear feature x hidden = MLP(h RoBERTa (x inpu t)). In this embodiment, the non-linear feature processor M uses an MLP (multi-layer perceptron). By concatenating the decision-making requirement and the preliminary decision-making solutions, the pre-trained model can fully explore the associated semantics between the two, and has a richer semantic representation compared to the pure preliminary decision-making solutions.

[0090] (2) Mixture-of-experts part

[0091] The mixture-of-experts part mainly consists of a sparse mixture of experts and multiple metric heads. The mixture-of-experts part consists of a sparse mixture of experts and multiple metric heads; the multiple metric heads include at least two metric heads, and each metric head uses a different evaluation mechanism to evaluate the same text to achieve feature complementarity. For an initial decision-making scheme, the perspective of evaluating its quality is multi-faceted. Using only a single evaluation metric, such as ROUGE-L, is not comprehensive. Using multiple metric heads for comprehensive evaluation can effectively avoid the biases caused by a single metric. At the same time, in order to adapt to diverse decision-making requirements and initial decision-making schemes, and enhance the robustness and generalization of the entire ranking method, a sparse mixture-of-experts system (MOE) is introduced.

[0092] For each metric head H, a set of sparse mixture-of-experts models E H ={e H,1 ,e H,2 ,…,e H,n} is introduced, and each set of models includes n experts in total; for each metric head, attention weights are calculated based on the intermediate feature representation x hidden . Each expert e H ={e H,1 ,e H,2 ,…,e H,n} in the expert set E H,i will obtain a weight α H,i , which is expressed as:

[0093]

[0094] where: score is the evaluation function in the attention mechanism, which is used to measure the correlation between x hidden and each expert e H,i , and dot product is used for evaluation scoring:

[0095]

[0096] where: is the weight matrix;

[0097] The outputs of the experts are weighted and summed to obtain the score s H of the metric head H:

[0098]

[0099] A loss function is constructed based on the scores of all metric heads:

[0100]

[0101] where: is the loss function; λ i is the weight of the metric head Hi; For measuring the measurement loss of the measurement head Hi; s i is the score of the measurement head Hi; y is the label score; p is the number of measurement heads;

[0102] Based on the weights of the measurement heads, the scores of all measurement heads are weighted and measured:

[0103]

[0104] where: s is the measurement score of each preliminary decision plan.

[0105] In this embodiment, the two metrics ROUGE-L and BARTScore are selected as the measurement heads H of the sorting method R 、H B 。ROUGE-L evaluates similarity through the longest common subsequence between texts, considering both word frequency and word order; while BARTScore is a scoring method based on the generative model BART, which adopts a different word matching evaluation mechanism from ROUGE, but directly evaluates the text quality using the generative model. These two metrics adopt different evaluation mechanisms and can achieve complementary features when evaluating the same text, improving the comprehensiveness and accuracy of the evaluation.

[0106] According to the selected metrics, the scores of the two measurement heads can be obtained as s ROUGE―L and s BARTScore 。Then the corresponding loss function can be expressed as:

[0107]

[0108] where is the loss function, λ ROUGE―L ,λ BARTScore are the weights of the two measurement heads respectively; is the measurement loss.

[0109] Finally, based on the weights λ ROUGE―L ,λ BARTScore of the above measurement heads, the scores s ROUGE―L ,s BARTScore of the two measurement heads are weighted and measured:

[0110]

[0111] where: s is the measurement score of each preliminary decision plan, and the preliminary decision plans are sorted according to this score.

[0112] 2.3. Adaptive Fusion Decision Strategy

[0113] The main purpose of AFDS is to adaptively select a suitable model as a fusion device based on the best candidate solutions to achieve the advantage fusion of the best solutions, thereby obtaining the final decision-making solution.

[0114] First, fuse the top K candidate decision-making solutions in the ranking. These excellent candidate decision-making solutions often have complementary advantages or show similar trends in certain aspects. By combining these solutions, not only can their respective advantages be retained, but also the defects of a single solution can be effectively weakened, thereby improving the accuracy and robustness of the overall decision-making solution. Reduce the possibility of deviation in the manufacturing decision-making solution.

[0115] Different from the re-design or selection of a new model proposed in some literatures, this embodiment adopts an adaptive fusion decision-making strategy and directly uses the fine-tuned LLM ranked first in the ranking method as the fusion device. This strategy aims to give full play to the effectiveness of existing excellent models. Specifically, the input of the fusion device is the decision-making requirement Q and the set of the top K candidate decision-making solutions {PD 1 , PD 2 , …, PD k} (for example, K = 2). To achieve the effective fusion of multiple candidate solutions, we sequentially connect the input and candidate solutions by using delimiters (such as "[candidate_id]") and reasonably optimize the instructions of the large model to guide the fusion device to extract and utilize the effective information in the best candidate solutions, thereby generating the final decision result RD.

[0116] The adaptive fusion decision-making mechanism not only improves the accuracy of decision-making, but also simplifies the complexity of model selection.

[0117] 3. Case Analysis

[0118] 3.1. Dataset

[0119] Taking the operation and maintenance of production equipment as the case study object, evaluate the effects of large language models and the LLM-MANUF integrated system in intelligent decision-making in manufacturing. The production equipment operation and maintenance dataset constructed in this embodiment is derived from the workshop equipment operation and maintenance records of an automobile whole vehicle manufacturing factory in Chongqing, China. The time span of this operation and maintenance record is 3 years, and the number of entries is 23k. The number of equipment involved in this operation and maintenance record is 113 types, and the number of components involved is 303 types. High-frequency failure components include equipment such as robots, sensors, and cylinders. Typical equipment failures are shown in Table 2; among them, robot failures include phenomena such as no movement of the robot and axis deviation errors, with a total of 2191 pieces, accounting for 9.25%; sensor failures are mainly manifested as the sensor not detecting the workpiece, with a total of 1237 pieces, accounting for 5.22%; cylinder failures include phenomena such as the cylinder not opening in place and the cylinder signal flashing off, with a total of 563 pieces, accounting for 2.38%.

[0120] Table 2 Typical Equipment Failures in the Whole Vehicle Manufacturing Process

[0121]

[0122] Since the workshop equipment operation and maintenance records are semi-structured data, they cannot be directly used by the system and need to be preprocessed, such as Figure 6 shown. Each column of the workshop equipment operation and maintenance records represents a kind of information, including but not limited to equipment name, main functional group, fault phenomenon and operation and maintenance measures. Among them, the main functional group refers to the main functional module of the equipment. For example, the VFG of the gripper robot is the gripper unit, the VFG of the spot welding robot is the welding tongs, and the VFG of the arc welding robot is the gun head. We process the above four types of information, use a general LLM to establish the relationship between components at the equipment-component level, the equipment and the fault phenomenon for the four types of information, and express all the information in a fluent form. Finally, a production equipment operation and maintenance data set is formed, which is constructed in the alpaca format, and some information can be seen in Table 3.

[0123] Table 3 Style of the Instruction Fine-tuning Data Set

[0124]

[0125] 3.2 Evaluation Metrics

[0126] The metrics adopted by the LLM-MANUF system in this embodiment are Bleu-4, Rouge-1, Rouge-2, and Rouge-L. BLEU-4 is a metric for measuring the quality of text generation, which is calculated through n-grams of 1 to 4 words to compare the similarity between the generated text and the reference text. The expression of BLEU is as follows:

[0127]

[0128] where: p n is the n-gram precision of the generated decision-making plan and the reference decision-making plan; ω n is the weight of the n-gram; N is the maximum n-gram length (in this embodiment, N = 4); BP is the length ratio of the generated decision-making plan and the reference decision-making plan.

[0129] Rouge-1, Rouge-2, and Rouge-L are metrics in the same series and are also used to measure text similarity. Rouge-1 and Rouge-2 are similar in metric calculation. The difference is that Rouge-1 only calculates the overlap degree of 1-gram (single word), while Rouge-2 calculates the overlap degree of 2-gram (two consecutive words). The expression of Rouge-N is as follows:

[0130]

[0131] Among them: {label} represents all tags, that is, the reference decision-making scheme; Count matc h (gram n ) represents the n-gram that exists in both the generated decision-making scheme and the reference decision-making scheme; Count(gram n ) represents all possible n-grams in the reference decision-making scheme.

[0132] ROUGE-L focuses on the matching of the longest common subsequence (LCS) between the generated text and the reference text, and its expression is as follows:

[0133]

[0134] Among them: RD is the generated decision-making scheme; m is the length of the reference decision-making scheme.

[0135] 3.3. Parameter Settings and Hardware

[0136] The fine-tuning of the large language model in this embodiment is all implemented through the llama_factory tool, and the decision-making sorting module is based on the pytorch framework. In this embodiment, LORA is used for fine-tuning. The maximum length cutoff_len of each input sample is 1024 tokens, the maximum number of samples max_samples used in training is 1000, the batch size batch_size for training on each device is 1, the gradient accumulation step is 8, the learning rate learning_rate is 0.0001, a cosine annealing learning rate scheduler is used, and the proportion of the warm-up phase is 10%, gradually increasing the learning rate.

[0137] All experiments in this embodiment are carried out on the server. The server operating system is Ubuntu 20.04.6 LTS, the CPU is AMD 5995WX, the memory is 8×32GB, and the GPU is 4×NVIDIA A6000 48GB.

[0138] 3.4. Experimental Results

[0139] In order to verify the effectiveness of the LLM-MANUF system proposed in this embodiment in integrating the advantages of multiple preliminary decision-making models, we conducted relevant experiments, and the specific results are shown in Table 4.

[0140] 3.4.1 The LLM-MANUF System Integrates the Advantages of All Parties

[0141] As can be seen from Table 4, after LORA fine-tuning, the 7B-level Llama3, Qwen2.5, Gemma2, Baichuan2, and Chatglm3 models selected in this embodiment all achieved good performance in the test set of the production equipment operation and maintenance dataset. Among them, the Qwen2.5 model had the highest score, with a BULE-4 score of 73.83, and the Chatglm3 model had the lowest score, with a BULE-4 of 66.87, and the difference between the two was 6.96. The LLM-MANUF system achieved the highest score in all indicators, with a BULE-4 score as high as 82.55, which was an improvement of 8.72 compared to Qwen2.5, the model with the highest score among the single fine-tuned models.

[0142] To further evaluate the performance of the system, full-parameter fine-tuning was introduced in the experiment. Generally speaking, full-parameter fine-tuning will have a certain degree of improvement in effect compared to LORA fine-tuning, and the specific results can also be learned from Table 4. In full-parameter fine-tuning, the performance of all models has improved compared to the LORA fine-tuning method; the Baichuan2 model replaced Qwen2.5 and became the best model, with a BULE-4 index value of 75.62, while the effect of Chatglm3 was still the worst. The full-parameter fine-tuning LLM replaced the LORA fine-tuning LLM to form the LLM-MANUF(full) system. Although the LLM-MANUF(full) system did not have as large an increase in score as the LLM-MANUF(LORA) system, its final score was obviously higher than that of the LORA-based integrated system, reaching 83.37.

[0143] 3.4.2 The LLM-MANUF System Challenges the High-Parameter Limit

[0144] Generally speaking, LLMs with larger parameter counts usually perform better in terms of performance than those with smaller parameter counts, but the limitation of the order of magnitude of parameter counts is not completely insurmountable. To prove the good effect of the LLM-MANUF system in breaking through the high-parameter limit, we introduced LLMs with parameter counts of 13B and 30B for experiments. Specifically, the Qwen2.5-14B-Instruct, Qwen2.5-32B-Instruct, and Llama2-Chinese-13b-Chat-ms models were considered. Since there is no 13B-level version of the llama3 model, we called the previous version llama2 of the same series. For the purpose of reducing the impact of version differences, the 7B-level llama2 was also involved in the experiment.

[0145] As can be seen from Table 4, the llama2-7B model performs worse than llama3 whether under the LORA fine-tuning method or full-parameter fine-tuning, while the llama2-13B model outperforms any llama3 model in terms of performance. This shows that even though llama2 is less performant than llama3 at the 7B level, the performance difference between model versions can be smoothed out by increasing the number of parameters, which also verifies the gap between different parameter levels of LLM.

[0146] Since the Qwen2.5 series covers all parameter levels from 7B to 30B, its performance comparison is more intuitive. As can be seen from Table 4, the Qwen2.5-14B and Qwen2.5-32B models have consistently maintained good performance. The BULE-4 score of Qwen2.5-14B is 78.05, and the BULE-4 score of Qwen2.5-32B is 83.46. And our LLM-MANUF system can easily outperform LLM at the 14B level, whether it is llama2 or Qwen2.5; for LLM at the 32B level, the LLM-MANUF system has an advantage in the two metrics of ROUGE-1 and ROUGE-L, and the other two metrics, BULE-4 and ROUGE-2, are also very close to QWEN2.5-32B. Generally speaking, our LLM-MANUF system has a good effect in breaking through the high-parameter limit.

[0147] 3.4.3 The LLM-MANUF system is an integrated whole

[0148] The LLM-MANUF system consists of multiple fine-tuned LLMs, sorting methods, and adaptive fusion decision-making strategies. Each component is indispensable to the LLM-MANUF system. In addition to evaluating the fine-tuned LLMs, it is also necessary to evaluate the sorting methods and fusion strategies. As can be learned from Table 4, when changing the sorting method to MLM

[36] and SimCLS, although the LLM-MANUF system can still maintain good accuracy, the improvement in its performance is not as good as that of adopting the DWMOE method, and the difference in BULE-4 can reach 2.96. Although the adaptive fusion decision-making strategy is only a small change, it can not only avoid the work of designing a new model but also improve the overall performance of the system. Fixing the fusion device, such as llama3, will inevitably lead to a decline in its performance. In addition, whether the llama model used as the fusion device is fine-tuned also affects the performance of the system. The performance of the original llama3 in the BULE-4 test is 1.8 lower than that of the fine-tuned llama3.

[0149] Table 4 Experimental Results

[0150]

[0151] Note: * indicates that the fusion device has not been fine-tuned.

[0152] 3.5 Analysis and Discussion

[0153] 3.5.1 Professional models are required for detailed decision-making solutions

[0154] To deeply analyze the differences among various models in intelligent decision-making in manufacturing, Table 5 shows the answers of each model in specific cases for visual presentation of details. In the table, the Chinese answers are shown on the left and the corresponding English translations are on the right. This is because in actual applications, engineers usually do not ask questions in English, so the English part in the experiment is not the direct output of the model.

[0155] In a specific case, the fault phenomenon is that the cylinder signal flashes intermittently. All major models can generate corresponding operation and maintenance decision-making solutions for this phenomenon, but there are differences in details compared with the solutions generated by the LLM-MANUF system. Whether it is ChatGPT-4o, Claude 3.5 or Qwen 2, the generated operation and maintenance decision-making solutions basically cover aspects such as cylinder air source, solenoid valve, signal line, control system, and wear interference. However, these solutions only provide macroscopic operation and maintenance suggestions and are of limited help in actual equipment operation and maintenance.

[0156] Each decision-maker in the LLM-MANUF system has been fine-tuned with a corpus in the manufacturing field and thus has more targeted professional knowledge. As can be seen from the table, for the fault phenomenon of intermittent flashing of the cylinder signal, the LLM-MANUF system provides more specific operation and maintenance measures, such as checking whether the cylinder locating pin is interfered, checking the position of the magnetic ring, and adjusting the inductor. Although these solutions have an inclusion relationship with the solutions proposed by general large models, in the actual operation process, the LLM-MANUF system proposed in this embodiment is more feasible.

[0157] Table 5 Operation and Maintenance Cases

[0158]

[0159]

[0160]

[0161] 3.5.2 The amount of contribution of decision-makers

[0162] To better demonstrate the contributions of each preliminary decision-making model to the LLM-MANUF system, we statistically analyzed the decision-making participation of each preliminary decision-making model in the test samples, that is, the proportion of the preliminary decision-making plans generated by the preliminary decision-making model that ranked in the top 2 in the ranking method. The specific results are shown in Table 6. We can see that the two models, Llama3 and Qwen2.5, have a high proportion of top 2 votes in the entire test sample, with the proportion exceeding 25% both in Rank1 and Rank2. In particular, the proportion of Qwen2.5 in Rank1 reaches 34.03%. Combining the results of the two fine-tuned models in Table 4, the better-performing single fine-tuned LLM also makes a greater contribution to the LLM-MANUF integrated system. Although the performance of the Chatglm3 model is not outstanding either in terms of its single performance or in the entire system, the voting proportion shows that it still helps to improve the performance of the LLM-MANUF system.

[0163] Table 6 Participation of Preliminary Decision-Makers in Integration

[0164]

[0165] 3.5.3 Better Prompts Facilitate Performance Improvement

[0166] During the fine-tuning process of large language models, prompts can guide the model to optimize for specific tasks. In vertical domains, such as the manufacturing domain, effective prompts can help the model focus on the knowledge and task requirements of the manufacturing industry. We adjusted the prompts in the constructed production equipment operation and maintenance dataset, and the adjusted prompts can be seen in Table 7. Subsequently, we fine-tuned each LLM using the dataset with adjusted prompts. The preliminary decision-making models after instruction adjustment have improved to varying degrees in various metrics of the decision-making task. For example, the BULE-4 of the LLAMA3 model has increased by 2.71 points, reaching 72.00. To more intuitively demonstrate the gain effect of prompt optimization on system performance, this embodiment selected several examples for analysis, and the details of specific cases are shown in Table 8. From the cases, it can be seen that better prompts can not only better learn the knowledge in the fine-tuning corpus but also more effectively stimulate the original pre-trained knowledge of the LLM, thereby enhancing the robustness of the LLM-MANUF system.

[0167] Regarding the LLM-MANUF system, the performance of the LLM-MANUF system improves correspondingly with the improvement of the performance of the preliminary decision-making model after fine-tuning, which is consistent with the law described in the previous experiments. The above experimental results show that in different manufacturing decisions, according to the requirements of the decision-making task and the characteristics of the data presented, different instructions can be designed so that the LLM can fully understand the task requirements and focus on and learn the knowledge required in the fine-tuning data.

[0168] Table 7 Adjusted Hints

[0169]

[0170] Table 8 Impact of Immediate Changes on the Decision-Making Plan

[0171]

[0172] 3.5.4 Quick feedback is essential

[0173] In the manufacturing decision-making process, the time required to generate decision-making plans is a key factor that must be taken seriously. We tested the average decision-making generation time of the LLM-MANUF system and fine-tuned LLMs, and the results are as Figure 7 shown. The time required for a single LLM to generate a decision-making plan shows an increasing trend as the number of its parameters increases. Specifically, when the parameter scale ranges from 7B to 14B, the increase in decision-making generation time is not significant, only increasing from about 18 seconds to 21 seconds; however, when the number of parameters increases to 30B, the decision-making generation time increases significantly to about 9 times that of the 7B-parameter-level LLM. It is worth noting that although the decision-making generation time of the LLM-MANUF system is about 6 times that of the 7B-parameter-level LLM, i.e., 111.28 seconds, compared with the decision-making generation time required by the 30B-parameter-level LLM, the LLM-MANUF system still shows certain timeliness advantages.

[0174] 4. Conclusions

[0175] In modern manufacturing, manufacturing decisions run through the entire life cycle of products from design to production. High-quality decisions can significantly improve production efficiency and productivity. However, as the complexity of manufacturing systems increases, existing manufacturing decision-making methods often struggle to effectively handle these complex correlation problems, which leads to manufacturing decision-making deviation problems. In addition, there are significant differences in the professional levels of manufacturing decision-making experts, which further exacerbates the occurrence of manufacturing decision-making deviations.

[0176] Regarding the problem of manufacturing decision-making bias, this embodiment proposes a fine-tuned large language model integration system for intelligent manufacturing decision-making, LLM-MANUF. The system aims to enhance the understanding and analysis capabilities of manufacturing decision-making requirements by integrating multiple LLMs fine-tuned in the manufacturing domain. With its powerful context semantic reasoning ability and pre-trained knowledge base, the LLM can deeply analyze manufacturing decision-making requirements and generate diverse decision-making schemes. This not only reduces the impact of traditional manufacturing corpora relying on expert experience on the decision-making process but also improves the objectivity and consistency of decision-making. LLM-MANUF introduces a dynamic weighted mixture of experts ranking method. By modeling the probability distribution of each candidate solution and under the guidance of multi-metric evaluation indicators, it effectively ranks the candidate solutions, thereby excluding options with a low correlation with actual decision-making requirements. Finally, based on the adaptive fusion decision-making strategy, LLM-MANUF can dynamically select the most suitable LLM as the fusion device according to the requirements of specific manufacturing scenarios, integrate the advantages of multiple best candidate solutions, and generate manufacturing decision-making schemes. Through the LLM-MANUF system, the accuracy and reliability of manufacturing decision-making schemes can be improved, and the decision-making bias caused by increased complexity and uneven expert levels can be effectively reduced. We selected the operation and maintenance of production equipment as the case study object and constructed a dataset for the operation and maintenance of production equipment for experiments. The experimental results show that the LLM-MANUF system we proposed can generate more accurate manufacturing decision-making schemes than single fine-tuned models and general large models, and it can achieve the effect of higher-parameter large language models with fewer computing resources.

[0177] The above-described embodiments are only preferred embodiments given to fully illustrate the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A fine-tuned large language model integration system for intelligent decision-making in manufacturing processes, characterized by: Including fine-tuning the large language model LLM, DWMOE module and AFDM module; The fine-tuned large language model LLM is set to multiple, each of the fine-tuned large language model LLM i Generate a preliminary decision plan PD based on the given decision requirement X i ; The DWMOE module defines a comprehensive ranking function Rank (PD i ), and based on multiple ranking metric functions {f1(i),f2(i),…,f m (i)} to perform weighted sorting to achieve the final sorting of all preliminary decision plans; m is the number of sorting metric functions; The AFDM module selects the top K preliminary decision solutions as candidate decision solutions through an adaptive fusion decision strategy and concatenates them to form a fusion instruction; The fine-tuned large language model LLM corresponding to the first-ranked preliminary decision plan is used as the fuser to generate the final manufacturing decision plan based on the fusion instructions.

2. The fine-tuning large language model integration system for intelligent decision-making in manufacturing process according to claim 1 is characterized by: The fine-tuned large language model LLM is obtained by using the LORA fine-tuning method based on the general domain large language model; the LORA fine-tuning method is: Freeze the weight matrix W of the general domain large language model and introduce a trainable low-rank matrix ΔW; the low-rank matrix ΔW uses two small matrices and The product of is: ΔW=AB T Where: r(A) = r(B) = r; Given a manufacturing decision requirement Q, the manufacturing decision requirement Q is represented as the input feature vector The feature vector x Q The feature mapping is performed through the linear transformation of a certain layer, and the fine-tuned weight matrix W′ is expressed as: W′=W+ΔW=W+AB T The linear transformation result h of manufacturing decision demand Q Q for: h Q =W′x Q =(W+AB T )x Q =Wx Q +A(B T x Q ) Where: A(B T x Q ) is the characteristic adjustment of manufacturing decision demand Q.

3. The fine-tuning large language model integration system for intelligent decision-making in manufacturing process according to claim 2 is characterized by: Based on different general-purpose large language models, the LORA fine-tuning method is used to obtain a plurality of different fine-tuned large language models {LLM1, LLM2, ..., LLM n }; Each fine-tuned large language model LLM i Combined with its own pre-trained general knowledge and manufacturing knowledge, it conducts in-depth analysis of the manufacturing decision demand Q, realizes the multi-dimensional semantic capture of manufacturing knowledge, and deeply understands the complex associations of the manufacturing system; at the same time, it fine-tunes the large language model LLM in parallel to perform multi-dimensional feature transformation and extraction in multiple feature subspaces, obtains relevant manufacturing decision information in different feature subspaces, and generates preliminary decision plans {PD1, PD2, …, PD n }; n is the number of fine-tuned large language models.

4. The fine-tuning large language model integration system for intelligent decision-making in manufacturing process according to claim 1 is characterized by: The DWMOE module includes a semantic parsing part and a hybrid expert part. The semantic parsing part is used to realize semantic feature extraction, and the hybrid expert part is used to realize probability distribution model construction and realize candidate sorting under the guidance of multi-metric heads.

5. The fine-tuning large language model integration system for intelligent decision-making in manufacturing process according to claim 4 is characterized by: The semantic parsing part includes a semantic parser S and a nonlinear feature processor M; the semantic parser S adopts a pre-training model that can effectively improve the quality of semantic features; the method for extracting semantic features in the semantic parsing part is: the decision requirement Q is combined with the preliminary decision plan {PD1, PD2, ..., PD n } to obtain the input representation x input =[Q;PD i ], the input represents x input Input into the semantic parser S; Extract the last layer feature representation h of the pre-trained model RoBERTa (x input ), and input it to the nonlinear feature processor M for nonlinear feature processing to obtain the nonlinear feature x hidden =MLP(h RoBERTa (x input )).

6. The fine-tuning large language model integration system for intelligent decision-making in manufacturing process according to claim 4 is characterized by: The hybrid expert part is composed of a sparse hybrid expert and a multi-metric head; the multi-metric head includes at least two metric heads, each of which uses a different evaluation mechanism to evaluate the same text to achieve feature complementarity; For each metric head H, a set of sparse mixture expert models E is introduced H ={e H,1 ,e H,2 ,…,e H,n }, each group of models includes n experts; for each metric head, based on the intermediate feature representation x hidden Calculate attention weight, expert set E H ={e H,1 ,e H,2 ,…,e H,n Each expert α H,i will get a weight α H,i , expressed as: Among them: score is the evaluation function in the attention mechanism, which is used to measure x hidden With each expert H,i The correlation is evaluated and scored using the dot product: in: is the weight matrix; The weighted sum of the expert's output is used to obtain the score s of the metric head H H : Construct the loss function based on the scores of all metric heads: in: is the loss function; i is the weight of the measurement head Hi; is the measurement loss of the measurement head Hi; s i is the score of the metric head Hi; y is the label score; p is the number of metric heads; Based on the weight of the metric head, the scores of all metric heads are weighted: Where: s is the metric score of each preliminary decision plan.

Images