A method, device and medium for mining association rules of aircraft assembly information based on fine-tuned large language model

Abstract

The present invention discloses a method, device, and medium for mining association rules of aircraft assembly information based on a fine-tuned large language model, relating to the field of electronic data digital processing. The method comprises the following steps: S1, collecting multi-source production factor data from aircraft assembly process documents; S2, performing domain adaptation based on a large language model fine-tuning method for low-rank adaptation; S3, extracting structured factor information and constructing a semantically optimized project set; S4, mining association rules between production factors and assembly links based on an association rule mining algorithm; and S5, optimizing and outputting a high-confidence association rule set. The present invention effectively improves the accuracy of information extraction and the depth of production factor association analysis by fine-tuning the large language model in combination with association rule mining, introducing a low-rank matrix, and deeply integrating the mining algorithm. This significantly improves the accuracy of identifying the association relationship between production factors, providing reliable data support and decision-making basis for optimizing the assembly process.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic data digital processing, in particular to information retrieval, and specifically to a method, device and medium for mining association rules of aircraft assembly information based on a fine-tuned large language model. Background Art

[0002] The aircraft final assembly manufacturing process involves a vast amount of production factor data, including complex information such as process steps, material requirements, tooling and equipment codes, and workstation locations. This data is typically stored in a variety of formats, including structured (such as AO process cards in MES systems), semi-structured (such as bills of materials in PDM systems), and unstructured (such as natural language descriptions of process instructions). These elements are interconnected during final assembly, significantly impacting assembly efficiency.

[0003] Traditional aircraft assembly information management and analysis methods rely primarily on manual experience and simple data analysis, making it difficult to comprehensively and accurately extract and analyze the relationships between these production factors. This poses numerous challenges to optimizing the assembly process and improving efficiency. These methods often suffer from technical issues such as inaccurate information extraction and incomplete relationship analysis when processing complex production factor data.

[0004] Therefore, a method that can efficiently and accurately mine the correlation between production factors in aircraft assembly process information is urgently needed. Summary of the Invention

[0005] The present invention aims to solve the problem of inaccurate extraction and association analysis of production factor information during aircraft assembly, overcomes the shortcomings of the existing technology, and provides a method, device and medium for mining aircraft assembly information association rules based on a fine-tuned large language model.

[0006] To achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a method for mining association rules of aircraft assembly information based on a fine-tuned large language model, comprising the following steps:

[0007] S1. Collect and pre-process multi-source production factor data in aircraft assembly process files;

[0008] S2. Fine-tuning a large language model based on low-rank adaptation for domain adaptation to enhance the ability to recognize specialized terminology.

[0009] S3. Extract structured element information through the fine-tuned large language model and build a semantically optimized item set;

[0010] S4. Mining association rules between production factors and final assembly based on association rule mining algorithm;

[0011] S5. Combine word vector clustering and semantic verification to optimize and output a high-confidence association rule set.

[0012] In a preferred embodiment of the present invention, in step S1, the multi-source production factor data includes structured and unstructured data, and includes the following sub-steps:

[0013] S11. Extract the AO number, workstation number, and station number from the structured data and perform standardized coding;

[0014] S12. Parse material codes and tooling equipment codes from semi-structured data;

[0015] S13. Extract professional terms, material codes, and tooling equipment codes from unstructured texts through natural language processing technology, and perform data cleaning based on regular expressions, domain dictionaries, and edit distance.

[0016] In a preferred embodiment of the present invention, in step S13, the unstructured text is extracted from the process steps described in natural language using a large language model, and the input format is defined as:

[0017] ;

[0018] in, It is process; It is a process AO number; It is a process Workstation number; It is a process Station number; It is a process Required materials, Is a collection of material numbers; It is a process Required tooling equipment, is a numbered set; It is a process A collection of key process information.

[0019] In a preferred embodiment of the present invention, the step S2 includes the following sub-steps:

[0020] S21. Insert a low-rank adapter module into the Transformer layer of the pre-trained large language model and adjust the model parameters through the low-rank matrix;

[0021] S22, freeze the original model parameters, train only the low-rank adapter module, and use the cross-entropy loss function and dynamic learning rate strategy for fine-tuning;

[0022] S23. Build a fine-tuning dataset based on domain-adapted prompt templates, and enhance information extraction accuracy by combining role definitions, task statements, and output examples.

[0023] In a preferred embodiment of the present invention, in the step S21, the low-rank matrix insertion is specifically as follows: inserting a matrix of rank next to the Query and Value matrices of each Transformer layer of the model. The low-rank adapter module, the weight update formula is:

[0024] ( , );

[0025] in, is the rank number; and is the row and column dimension of the original model weight matrix; is a set of real numbers; Transformer is the model architecture, and the Transformer layer contains a self-attention mechanism and a feedforward neural network to capture contextual dependencies in the text; Query is the query matrix in the self-attention mechanism, which is used to calculate the correlation weights between different positions in the input sequence; Value is the value matrix in the self-attention mechanism, which is used to apply weights to the original input to generate the final context-aware representation.

[0026] In a preferred embodiment of the present invention, the step S3 includes the following sub-steps:

[0027] S31. Use a multi-agent mechanism to verify the format of the extraction results and match them with domain terms, and use a clustering algorithm to merge synonym classes and eliminate meaningless words;

[0028] The multi-agent mechanism includes: a first agent module: verifying the format standardization of the extraction results; a second agent module: verifying the compliance of terms and codes based on the aviation manufacturing field terminology library;

[0029] S32. Encode the clustering results into standard labels to generate semantically consistent structured item sets.

[0030] In a preferred embodiment of the present invention, in step S4, the association rules between the production factors and the final assembly process are based on the FP-Growth algorithm, specifically including: frequent item set mining and generating association rules based on the frequent item sets.

[0031] In a preferred embodiment of the present invention, the frequent itemset mining is as follows: the frequency of occurrence of each itemset in the transaction database is counted, and for the itemset ( is the set of all items, all items are all different single basic items in the data set), and its support is calculated as:

[0032] , ;

[0033] in, is the total number of transactions; is the indicator function, when 1 when it is, otherwise 0; It is an item set Throughout the transaction database The number of times it appears in It is a transactional database A single transaction in , that is, a collection of data items;

[0034] Keep support ≥ minimum threshold Itemsets, collections of frequent itemsets Defined as:

[0035] ;

[0036] Create an index for each frequent item, record its position and global frequency in the FP tree, and the head pointer table Is a dictionary with keys as frequent items , the value is:

[0037] ;

[0038] in, This item Support count of ; Points to all the items in the FP tree. Node list;

[0039] Compress the transaction database into a prefix tree structure, where each node Contains the following properties:

[0040] ;

[0041] in, is the item represented by the current node; is the number of occurrences of paths from the root to the node; Is the parent node pointer; is a dictionary of child nodes;

[0042] Build rules: For each transaction , sorted in descending order by frequent item support and inserted into the tree, and the paths sharing the same prefix are counted together;

[0043] Conditional pattern base generation: For item , its conditional pattern is based on is all the numbers in the FP tree that are The path collection at the end, each path is in the form of:

[0044] ;

[0045] Recursive formula: For each term , recursively construct the conditional FP tree , and mine its frequent itemsets:

[0046] ;

[0047] in, In the recursive process of FP-Growth algorithm, it means that the current item The set of new frequent itemsets generated by the conditional pattern base; It is in the Conditional pattern base Rerun the FP-Growth algorithm on the same minimum support threshold To mine frequent itemsets;

[0048] Finally merge all Get global frequent itemsets;

[0049] The association rule calculation: generating rules based on frequent item sets, and screening high-value rules by confidence and lift;

[0050] The confidence level indicates that The transaction also includes the post-processing The proportion of transactions, that is, the accuracy of the rules:

[0051] ;

[0052] in, is included and The number of transactions; is included The number of transactions;

[0053] Lift and The higher the correlation strength between and The stronger the association relationship, if the lift value is greater than 1, the association rule is considered valuable:

[0054] ;

[0055] in, It is a measurement rule The strength of the association, reflecting and Is the co-occurrence of statistic significant? It's a rule confidence level; It is an item set support.

[0056] The present invention provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor; the memory stores a computer program executable by the processor, and the computer program is executed by the processor so that the processor can execute any one of the above-mentioned aircraft assembly information association rule mining methods based on a fine-tuned large language model.

[0057] The present invention provides a computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a processor to implement the aircraft assembly information association rule mining method based on a fine-tuned large language model as described above when executed.

[0058] The present invention solves the defects existing in the background technology and has the following beneficial effects:

[0059] (1) The present invention provides a method, device and medium for mining association rules of aircraft assembly information based on fine-tuning a large language model. By fine-tuning the large language model in combination with association rule mining, the model's ability to recognize and extract information in professional terms in the aviation field is significantly enhanced. The introduction of low-rank matrices allows the model fine-tuning process to only require adjusting a small number of parameters, which not only maintains the general knowledge of the pre-trained model but also improves the accuracy of aviation professional tasks, effectively reducing training complexity and resource consumption. The in-depth combination of mining algorithms effectively improves the accuracy of information extraction and the depth of production factor association analysis, thereby greatly improving the recognition accuracy of the association relationship between production factors and providing reliable data support and decision-making basis for the optimization of the assembly process.

[0060] (2) In the present invention, low-rank adapter technology is adopted in the model fine-tuning stage. By inserting a low-rank matrix into the Transformer layer of the pre-trained model and freezing the original parameters, efficient domain adaptation of the large language model is achieved. The lightweight low-rank matrix greatly reduces the scale of parameter adjustment, enabling the model to converge quickly in the aviation professional terminology recognition task, while avoiding the excessive dependence of full parameter fine-tuning on computing resources. The combination of dynamic learning rate strategy and cross-entropy loss function further improves the stability and convergence speed of the fine-tuning process, significantly enhancing the adaptability of the model in the aviation field and significantly reducing the training cost.

[0061] (3) The association rule mining in the present invention adopts the FP-Growth algorithm combined with a multi-index screening mechanism. Through frequent item set mining and promotion screening, it not only ensures the high-frequency co-occurrence of rules, but also quantifies the actual correlation strength between the antecedent and the consequent, avoiding the false correlation caused by the high confidence trap in traditional methods. The combination of multi-dimensional evaluation rules ensures the actual business correlation strength of the rules and enhances the interpretability of the rules, thereby improving the mining accuracy of association rules and providing an accurate basis for workstation resource allocation and material scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments described in the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive efforts.

[0063] Figure 1 This is a flow chart of a method for mining association rules of aircraft assembly information based on a fine-tuned large language model according to a preferred embodiment of the present invention;

[0064] Figure 2 2 is a data diagram of the ChatGLM model fine-tuning process according to a preferred embodiment of the present invention;

[0065] Figure 3 is a schematic diagram of the extraction structure of a preferred embodiment of the present invention;

[0066] Figure 4 Schematic diagram of association rules according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0067] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0068] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0069] like Figure 1 As shown, a method for mining association rules of aircraft assembly information based on a fine-tuned large language model includes the following steps:

[0070] S1. Collect and pre-process multi-source production factor data in aircraft assembly process files;

[0071] S2. Fine-tuning a large language model based on low-rank adaptation for domain adaptation to enhance the ability to recognize specialized terminology.

[0072] S3. Extract structured element information through the fine-tuned large language model and build a semantically optimized item set;

[0073] S4. Mining association rules between production factors and final assembly based on association rule mining algorithm;

[0074] S5. Combine word vector clustering and semantic verification to optimize and output a high-confidence association rule set.

[0075] In this embodiment, in step S1, the multi-source production factor data includes structured and unstructured data, including the following sub-steps:

[0076] S11. Extract the AO number, workstation number, and station number from the structured data and perform standardized coding;

[0077] S12. Parse material codes and tooling equipment codes from semi-structured data;

[0078] S13. Extract professional terms, material codes, and tooling equipment codes from unstructured texts through natural language processing technology, and perform data cleaning based on regular expressions, domain dictionaries, and edit distance.

[0079] In a specific embodiment, in step S11, the structured data source comes from the AO process card exported from the MES system, wherein the standardized coding rule is specifically as follows:

[0080] AO number: The format is "AO-XXXXX" (e.g. AO-5720C41073), which must be a combination of the prefix "AO" and 10 alphanumeric characters.

[0081] Work station number (WS): The format is "number + letter" (e.g. 200B), and must be 3 characters long. Zeros are added to missing digits (e.g. 200B → 200B, no zeros are added).

[0082] Station number (ST): The format is "XXX" (such as 101), which must be a 3-digit number with zeros added to missing digits (such as 1→001).

[0083] In a specific embodiment, in step S12, the semi-structured data source comes from the BOM list of the PDM system, and the parsing rules are specifically as follows:

[0084] Material code: follow the "MAT-XXXXX" format (such as MAT-3205A);

[0085] Tool equipment code: follow the "TOOL-XXXX" format (such as TOOL-8892).

[0086] In a specific embodiment, in step S13, unstructured text is extracted from the process steps described in natural language using a large language model (ChatGLM3-6B). The input format is defined as:

[0087] ;

[0088] in, It is process; It is a process AO number; It is a process Workstation number; It is a process Station number; It is a process Required materials, Is a collection of material numbers; It is a process Required tooling equipment, is a numbered set; It is a process A collection of key process information.

[0089] Furthermore, the text preprocessing process adopts a multi-stage cleaning strategy:

[0090] Regular expression filtering: extract standard codes such as "XXX-XXXX";

[0091] Domain dictionary matching: load domain-specific vocabulary for precise matching;

[0092] Correct spelling errors based on edit distance.

[0093] In this embodiment, the step S2 includes the following sub-steps:

[0094] S21. Insert a low-rank adapter module into the Transformer layer of the pre-trained large language model and adjust the model parameters through the low-rank matrix;

[0095] S22, freeze the original model parameters, train only the low-rank adapter module, and use the cross-entropy loss function and dynamic learning rate strategy for fine-tuning;

[0096] S23. Build a fine-tuning dataset based on domain-adapted prompt templates, and enhance information extraction accuracy by combining role definitions, task statements, and output examples.

[0097] In a specific embodiment, in step S21, the low-rank matrix insertion is specifically as follows: inserting a low-rank matrix of rank next to the Query and Value matrices of each Transformer layer of the model (ChatGLM3-6B) The low-rank adapter module, the weight update formula is:

[0098] ( , );

[0099] in, is the rank number ( ); and is the row and column dimension of the original model weight matrix; is a set of real numbers; Transformer is the model architecture, and the Transformer layer contains a self-attention mechanism and a feedforward neural network to capture contextual dependencies in the text; Query is the query matrix in the self-attention mechanism, which is used to calculate the correlation weights between different positions in the input sequence; Value is the value matrix in the self-attention mechanism, which is used to apply weights to the original input to generate the final context-aware representation.

[0100] Example: Assume that the query matrix dimension of a Transformer layer is 768×768, then the low-rank matrix The dimensions are 768×8, is 8×768, after adjustment The dimensions are still 768×768.

[0101] In a specific embodiment, in step S22, as Figure 2 As shown, the hidden state of the input is represented as ,in Is the feature dimension, freezing the original model parameters is specifically: during the fine-tuning process, only the low-rank matrix is updated and Parameters, where The parameters of the normal distribution initialization, is the variance, which indicates the degree of dispersion of the data. Initially 0, original model parameters Keep it fixed;

[0102] For the input sequence and tags , the cross entropy loss function is:

[0103] ;

[0104] in, is the classification category number (such as aviation terminology, material code, etc.); The model predicts class probability; is the one-hot encoding of the true label;

[0105] Example: If the model needs to extract the term "wing-body docking" and the code "MAT-3205A" from the text "wing-body docking requires MAT-3205A", then the label is the corresponding one-hot encoded vector, and the loss function measures the difference between the predicted distribution and the true label;

[0106] Dynamically adjust the learning rate according to the gradient variance of the low-rank adapter module to prevent gradient explosion or oscillation: After each training round, calculate the gradient variance of all low-rank adapter parameters:

[0107] ;

[0108] in, is the total number of low-rank adapter parameters; is a low-rank adapter module (i.e., a low-rank matrix inserted and ) in a parameter; is the mean gradient;

[0109] Learning rate adjustment rule: If , the learning rate decays to 50% of the original value:

[0110] ;

[0111] Adopt Adam optimizer and set the learning rate to , the batch size is 32, and the training rounds are 10 until the loss function converges.

[0112] In a specific embodiment, in step S23, a prompt template is designed based on the CRISPE framework, and a fine-tuning dataset is constructed in combination with a few-shot example. The example template is as follows:

[0113] “Technical terms”: [“wing-body docking”, “fuel line installation”],

[0114] “Material Code”: [“MAT-3205A”, “MAT-1876B”],

[0115] “Tooling equipment code”: [“TOOL-8892”].

[0116] In this embodiment, the step S3 includes the following sub-steps:

[0117] S31. Use a multi-agent mechanism to verify the format of the extraction results and match them with domain terms, and use a clustering algorithm to merge synonym classes and eliminate meaningless words;

[0118] S32. Encode the clustering results into standard labels to generate semantically consistent structured item sets.

[0119] In a specific embodiment, in step S31, the extraction result is as follows: Figure 3 As shown; the multi-agent mechanism includes:

[0120] The first agent module: verifies the format standardization of the extraction results;

[0121] The second agent module: verifies the compliance of terms and codes based on the terminology database in the aviation manufacturing field.

[0122] Further, the input example of the first agent module is:

[0123] “Professional terms”: [“wing-body docking”, “fuel line installation”, “name”],

[0124] “Material Code”: [“MAT-3205A”, “MAT-32O5A”, “MAT-1876B”],

[0125] “Tooling equipment code”: [“TOOL-8892”, “TOOL-8892”].

[0126] First-agent processing: Eliminate invalid items: "name" (non-term), "MAT-32O5A" (including the letter O), "TOOL-8892" (including Chinese characters).

[0127] Output:

[0128] “Technical terms”: [“wing-body docking”, “fuel line installation”],

[0129] “Material Code”: [“MAT-3205A”, “MAT-1876B”],

[0130] “Tooling equipment code”: [“TOOL-8892”].

[0131] Verification term similarity matching rules of the second agent module:

[0132] Exact match: directly compare whether the term / code exists in the domain library;

[0133] Fuzzy matching: For terms that are not exactly matched, calculate the cosine similarity with the terms in the domain library and retain those with similarity ≥ 0.8;

[0134] Cosine similarity: ;

[0135] in, and is the word vector of the two words (such as the 300-dimensional vector generated by Word2Vec); is the vector dimension (300).

[0136] Furthermore, the K-means clustering algorithm is used to group word vectors. By presetting the number of clusters (e.g., K=10), the algorithm automatically groups semantically similar words into the same cluster. For example, "connect" and "interconnect" might be grouped together, while "tighten" might be grouped separately. During the clustering process, the system continuously optimizes the distribution of words within a cluster until the cluster center no longer changes significantly.

[0137] In a specific embodiment, in step S32, the clustering results are based on professional dictionaries or expert knowledge in the field of aviation manufacturing, and semantic labels are assigned to each cluster. For example, a cluster containing "connection" and "interconnection" is defined as a "connection class," and a cluster containing "fastening" and "bolt" is defined as a "fastener class." Automated processing is performed using the following rules:

[0138] Domain dictionary matching: Count the categories of words in the cluster in the domain dictionary and select the category with the highest frequency as the label.

[0139] Label encoding rules: Assign a unique code (such as C1-C10) to each semantic category and establish a mapping table, as shown in Table 1:

[0140] Table 1: Semantic category mapping table

[0141]

[0142] In this embodiment, in step S4, the association rules between production factors and final assembly are based on the FP-Growth algorithm, which specifically includes: frequent item set mining and generating association rules based on the frequent item sets.

[0143] Frequent itemset mining:

[0144] Count the frequency of occurrence of each item set in the transaction database. ( is the set of all items, which are all different single basic items in the data set, such as workstations, materials, and terms. Its support is calculated as:

[0145] , ;

[0146] in, is the total number of transactions; is the indicator function, when 1 when it is, otherwise 0; It is an item set Throughout the transaction database The number of times it appears in It is a transactional database A single transaction in , that is, a collection of data items;

[0147] Keep support ≥ minimum threshold Itemsets, collections of frequent itemsets Defined as:

[0148] ;

[0149] Create an index for each frequent item, record its position and global frequency in the FP tree, and the head pointer table Is a dictionary with keys as frequent items , the value is:

[0150] ;

[0151] in, This item Support count of ; Points to all the items in the FP tree. Node list;

[0152] Compress the transaction database into a prefix tree structure, where each node Contains the following properties:

[0153] ;

[0154] in, is the item represented by the current node; is the number of occurrences of paths from the root to the node; Is the parent node pointer; is a dictionary of child nodes (with items as keys);

[0155] Build rules: For each transaction , sorted in descending order by frequent item support and inserted into the tree, and the paths sharing the same prefix are counted together;

[0156] Conditional pattern base generation: For item , its conditional pattern is based on is all the numbers in the FP tree that are The path collection at the end, each path is in the form of:

[0157] ;

[0158] Recursive formula: For each term , recursively construct the conditional FP tree , and mine its frequent itemsets:

[0159] ;

[0160] in, In the recursive process of FP-Growth algorithm, it means that the current item The set of new frequent itemsets generated by the conditional pattern base; It is in the Conditional pattern base Rerun the FP-Growth algorithm on the same minimum support threshold To mine frequent itemsets;

[0161] Finally merge all Get the global frequent itemsets.

[0162] Association rule calculation: Generate rules based on frequent item sets and filter high-value rules by confidence and lift;

[0163] Confidence is expressed in terms of the antecedents (combination of production factors) The affairs also include post-assembly workstations or key processes The proportion of transactions, that is, the accuracy of the rules:

[0164] ;

[0165] in, is included and The number of transactions; is included The number of transactions;

[0166] Lift and The strength of the association between:

[0167] ;

[0168] in, It is a measurement rule The strength of the association, reflecting and Is the co-occurrence of statistic significant? It's a rule confidence level; It is an item set support level;

[0169] Specifically, if the lift value is greater than 1, the association rule is considered valuable:

[0170] Lift = 1: and Independent, unrelated;

[0171] Lift > 1: and Positive correlation, The appearance of Probability of occurrence;

[0172] Lift < 1: and Negative correlation, The appearance of Probability of occurrence.

[0173] It should be noted that in the aircraft assembly scenario, if the lift of a rule (such as {workstation 200B} ⇒ {material MS21043-3}) is greater than 1, it means that the probability of using this material at this workstation is significantly higher than the global distribution probability of this material, indicating a strong business correlation between the two, rather than random coincidence.

[0174] It is understandable that the proposed algorithm can avoid the high confidence trap and neglect of independence test problems of traditional association rule mining methods (such as Apriori and FP-Growth) that usually rely only on support and confidence to filter rules:

[0175] High confidence trap: If the consequent ( ) itself has a high degree of support (such as a material is widely used in all workstations), even if and No matter, rules The confidence level of may also be high, leading to spurious associations;

[0176] Ignoring independence tests: Unquantified and The actual correlation strength may be misjudged as accidental co-occurrence as a valid rule.

[0177] Example:

[0178] For example, if a process at workstation 200B requires universal screws 90% of the time (support = 30%), then the confidence level of the rule {workstation 200B} ⇒ {universal screws} is 90%, which appears to be a strong association. However, if the lift is 1 (i.e., the average usage rate of universal screws across all workstations is also 30%), this association is meaningless, simply because universal screws are frequently used.

[0179] In this embodiment, the step S5 includes the following sub-steps:

[0180] S51, performing secondary clustering on the optimized structured item set to eliminate noise data;

[0181] S52. Align the structured item set with the association rule set, generate a uniformly expressed rule set, and output a final high-confidence rule set.

[0182] In a specific implementation, in step S51, the Word2Vec model is first used to generate a 300-dimensional word vector for the optimized Chinese word segmentation result in step S32 to capture the semantic similarity between terms, and then the K-means clustering algorithm same as step S31 is applied to perform unsupervised clustering on the word vector, setting the number of clusters K=10, and iterating until the intra-cluster distance converges.

[0183] In a specific embodiment, in step S52, the structured item set is aligned with the association rule set, and a semantically unified and streamlined rule set is generated through mapping, merging and recalculation;

[0184] Update label encoding through rule item mapping: Based on the secondary clustering result of step S51, record the attribution relationship between the old label and the new label, replace the original items in the association rule (such as "C1" and "C2") with the new labels after secondary clustering (such as "C1-2"), and establish a label mapping table; traverse the antecedent of each rule ( ) and the consequent ( ), replace the old label according to the mapping table;

[0185] Eliminate redundant rules by merging rules: If the antecedents and consequents of multiple rules are completely consistent, they are merged into one rule, the support count is accumulated, and the support, confidence, and lift indicators are recalculated;

[0186] Delete rules whose indicators do not meet the standards due to label merging or data updates. The filtering conditions are:

[0187] Support: below the threshold (0.03);

[0188] Confidence: below threshold (0.85);

[0189] Lift: ≤1 (no real relevance);

[0190] Thus, the rules with confidence ≥ 0.85 and lift > 1 are output to guide production scheduling and resource optimization of final assembly.

[0191] Example:

[0192] like Figure 4 As shown in the figure, the confidence of the merged rule {C1-2} ⇒ {workstation 200B} is 1.0, and the lift is 1.0 / 0.02 = 50 (assuming the global support of workstation 200B is 0.02). This rule is retained. If the lift of a rule is 0.8 (<1), it is removed.

[0193] The present invention provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor; the memory stores a computer program executable by the processor, and the computer program is executed by the processor so that the processor can execute the above-mentioned aircraft assembly information association rule mining method based on a fine-tuned large language model.

[0194] The present invention provides a computer-readable storage medium storing computer instructions, which are used to enable a processor to implement the above-mentioned aircraft assembly information association rule mining method based on a fine-tuned large language model when executing the computer instructions.

[0195] In summary, the present invention significantly enhances the model's recognition and information extraction capabilities in aviation professional terminology by fine-tuning the large language model in combination with association rule mining. The introduction of low-rank matrices requires only a small number of parameters to be adjusted in the model fine-tuning process, which not only maintains the general knowledge of the pre-trained model, but also improves the accuracy of aviation professional tasks, effectively reducing training complexity and resource consumption. The in-depth combination of mining algorithms effectively improves the accuracy of information extraction and the depth of production factor association analysis, thereby greatly improving the recognition accuracy of the association relationship between production factors, and providing reliable data support and decision-making basis for the optimization of the final assembly process.

[0196] The above description is based on the ideal embodiment of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be considered as illustrative and non-restrictive in all respects. The scope of the present invention is defined by the appended claims, not the foregoing description, and it is intended that all variations within the meaning and range of equivalents of the claims be encompassed within the present invention. Any reference signs in the claims should not be construed as limiting the claim to which they relate.

[0197] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for mining association rules of aircraft assembly information based on a fine-tuned large language model, characterized by: The following steps are involved: S1. Collect and pre-process multi-source production factor data in aircraft assembly process files; S2. Fine-tuning a large language model based on low-rank adaptation for domain adaptation to enhance the ability to recognize specialized terminology. S3. Extract structured element information through the fine-tuned large language model and build a semantically optimized item set; S4. Mining association rules between production factors and final assembly based on association rule mining algorithm; S5. Combine word vector clustering and semantic verification to optimize and output association rule sets with a confidence level ≥ 0.85; In the step S4, the association rules between the production factors and the final assembly process are based on the FP-Growth algorithm, specifically including: frequent item set mining and generating association rules based on the frequent item sets; Frequent item set mining: statistics the frequency of occurrence of each item set in the transaction database, , is the set of all items, all items are all different single basic items in the data set, and their support is calculated as: ， ； in, is the total number of transactions; when 1 when it is, otherwise 0; The item set in the entire transaction database The number of times it appears in the transaction database is a single transaction, that is, a set of data items; Retain itemsets with support ≥ minimum threshold, the set of frequent itemsets Defined as: ； Create an index for each frequent item, record its position and global frequency in the FP tree, and the head pointer table Is a dictionary with keys as frequent items , the value is: ； in, This item Support count of ; Points to all the items in the FP tree. Node list; Compress the transaction database into a prefix tree structure, where each node Contains the following properties: ； in, is the item represented by the current node; is the number of occurrences of paths from the root to the node; Is the parent node pointer; is a dictionary of child nodes; Build rules: For each transaction , sorted in descending order by frequent item support and inserted into the tree, and the paths sharing the same prefix are counted together; Conditional pattern base generation: For item , its conditional pattern is based on is all the numbers in the FP tree that are The path collection at the end, each path is in the form of: ； Recursive formula: For each term , recursively construct the conditional FP tree , and mine its frequent itemsets: ； in, In the recursive process of FP-Growth algorithm, it means that the current item The set of new frequent itemsets generated by the conditional pattern base; It is in the Conditional pattern base Rerun the FP-Growth algorithm on the same minimum support threshold To mine frequent itemsets; Finally merge all Get global frequent itemsets; The association rule calculation: generating rules based on frequent item sets, and screening high-value rules by confidence and lift; The confidence level indicates that The transaction also includes the post-processing The proportion of transactions, that is, the accuracy of the rules: ； in, is included and The number of transactions; is included The number of transactions; Lift and The higher the correlation strength between and The stronger the association relationship, if the lift value is greater than 1, the association rule is considered valuable: ； in, It is a measurement rule The strength of the association, reflecting and Is the co-occurrence of statistic significant? It's a rule confidence level; It is an item set support.

2. The method for mining association rules of aircraft assembly information based on a fine-tuned large language model according to claim 1, characterized in that: In step S1, the multi-source production factor data includes structured and unstructured data, including the following sub-steps: S11. Extract the AO number, workstation number, and station number from the structured data and perform standardized coding; S12. Parse material codes and tooling equipment codes from semi-structured data; S13. Extract professional terms, material codes, and tooling equipment codes from unstructured texts through natural language processing technology, and perform data cleaning based on regular expressions, domain dictionaries, and edit distance.

3. The method for mining association rules of aircraft assembly information based on a fine-tuned large language model according to claim 2, characterized in that: In the step S13, the unstructured text is extracted from the process steps described in natural language using a large language model, and the input format is defined as: ； in, It is process; It is a process AO number; It is a process Workstation number; It is a process Station number; It is a process Required materials, Is a collection of material numbers; It is a process Required tooling equipment, is a numbered set; It is a process A collection of key process information.

4. The method for mining association rules of aircraft assembly information based on a fine-tuned large language model according to claim 1, characterized in that: The step S2 includes the following sub-steps: S21. Insert a low-rank adapter module into the Transformer layer of the pre-trained large language model and adjust the model parameters through the low-rank matrix; The Transformer is a model architecture, and the Transformer layer includes a self-attention mechanism and a feedforward neural network to capture contextual dependencies in text; S22, freeze the original model parameters, train only the low-rank adapter module, and use the cross-entropy loss function and dynamic learning rate strategy for fine-tuning; S23. Build a fine-tuning dataset based on domain-adapted prompt templates, and enhance information extraction accuracy by combining role definitions, task statements, and output examples.

5. The method for mining association rules of aircraft assembly information based on a fine-tuned large language model according to claim 4, characterized in that: In the step S21, the low-rank matrix insertion is specifically as follows: inserting a low-rank matrix of rank next to the Query and Value matrices of each Transformer layer of the model. The low-rank adapter module, the weight update formula is: ， ， ； in, is the rank number; and is the row and column dimension of the original model weight matrix; is a set of real numbers; Transformer is the model architecture, and the Transformer layer contains a self-attention mechanism and a feedforward neural network to capture contextual dependencies in the text; Query is the query matrix in the self-attention mechanism, which is used to calculate the correlation weights between different positions in the input sequence; Value is the value matrix in the self-attention mechanism, which is used to apply weights to the original input to generate the final context-aware representation.

6. The method for mining association rules of aircraft assembly information based on a fine-tuned large language model according to claim 1, characterized in that: The step S3 includes the following sub-steps: S31. Use a multi-agent mechanism to verify the format of the extraction results and match them with domain terms, and use a clustering algorithm to merge synonym classes and eliminate meaningless words; The multi-agent mechanism includes: a first agent module: verifying the format standardization of the extraction results; a second agent module: verifying the compliance of terms and codes based on the aviation manufacturing field terminology library; S32. Encode the clustering results into standard labels to generate semantically consistent structured item sets.

7. An electronic device, characterized in that: The electronic device includes: at least one processor, and a memory communicatively connected to the at least one processor; the memory stores a computer program executed by the processor, and the computer program is executed by the processor so that the processor can execute the aircraft assembly information association rule mining method based on a fine-tuned large language model according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the aircraft assembly information association rule mining method based on a fine-tuned large language model according to any one of claims 1 to 6 when executed.

Images