Aviation product quality report verification method and system based on pre-training language model

Abstract

The invention relates to the technical field of aviation product quality report key data extraction and verification, in particular to an aviation product quality report verification method and system based on a pre-training language model, and the method comprises the steps: S1, obtaining an aviation product quality return-to-zero report text, extracting a feature length, and carrying out the merging and partitioning operation; s2, extracting a key information feature text of the aviation product quality return-to-zero report text, and generating an inspection prompt; s3, performing context positioning on paragraphs of the aviation product quality return-to-zero report text; and S4, checking the aviation product quality return-to-zero report text by using a large-scale language model. According to the method, the quality problem return-to-zero report is analyzed by using the large-scale pre-training language model, so that the extraction accuracy of key information is improved, the report content is automatically checked, and the integrity and logic consistency of the report content are ensured; according to the method, common problems in quality control can be found, the improvement scheme is optimized, and reliable data support is provided for improvement and standardization of a quality return-to-zero report.

Description

Technical Field

[0001] The present invention relates to the technical field of extracting and verifying key data of aviation product quality reports, and in particular to a method and a system for verifying aviation product quality reports based on a pre-trained language model. Background Art

[0002] Zero report plays an important role in quality control, and its review and verification process is the key link to ensure that product quality problems are effectively solved. Zero report usually contains a lot of technical details, involving complex field terms and multi-level logical relationships. It must clearly describe the time, location, phenomenon and environmental conditions of the quality problem, and conduct a detailed analysis of the nature of the problem, product structure, design history, root cause, etc. Therefore, the inspection and verification process of the report has high requirements for the integrity, consistency and logical rigor of the information.

[0003] In practical applications, the inspection and verification tasks of zeroing reports are often time-consuming and require high professional knowledge and experience of personnel. Common problems during the inspection process include incomplete information extraction, inconsistent content, and logical errors in complex scenarios, all of which may affect the zeroing effect of quality issues. The present invention aims to improve the accuracy of key information extraction of zeroing reports through automated means, ensure the comprehensiveness and logical consistency of information, and then realize the automation of the report verification process, thereby improving the efficiency and consistency of the quality control process.

[0004] With the development of large-scale pre-trained language models (LLMs), text analysis technology based on natural language processing (NLP) has gradually been applied in the field of quality control and zero-reporting. In the existing technology, encoder-only, decoder-only, encoder-decoder and other architecture models are widely used for information extraction and text generation. However, these models still have problems such as incomplete information extraction and inconsistent generated content when processing zero-report analysis.

[0005] At present, typical applications of large-scale language models include Encoder-Only architecture models such as BERT (Bidirectional Encoder Representations from Transformers), which are mainly used for text classification and information extraction; Decoder-Only architecture models such as GPT (Generative Pre-trained Transformer) are used for long text generation. Although these models have demonstrated excellent performance, there are many challenges in specific applications of quality control, such as content verification and automatic generation of zero reports. The present invention aims to use large-scale language models to improve the accuracy of key data extraction of zero reports and realize automatic verification and report generation. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides an aviation product quality report verification method based on a pre-trained language model. The large-scale pre-trained language model is used to analyze the quality problem zeroing report, improve the accuracy of key information extraction, and realize automatic verification of the report content to ensure its integrity and logical consistency. The present invention can discover common problems in quality control and optimize improvement plans, providing reliable data support for the improvement and standardization of quality zeroing reports.

[0007] To achieve the above object, the present invention discloses an aviation product quality report verification method based on a pre-trained language model, which comprises:

[0008] S1: Obtain the aviation product quality zeroing report text, extract the feature length, and perform merging and block operations;

[0009] S11: Determine the mass zeroing report text and the maximum input text length; set the mass zeroing report text D to contain n text paragraphs {P1,…P i …,P n}, and the maximum input text length of the model is L max ,limit the limit of text blocks;

[0010] S12: Use the language model to calculate each text paragraph P i The characteristic length len(P i )for:

[0011] len(P i )=len(Tokenizer(P i ));

[0012] Among them, len(P i ) is the i-th text paragraph P i The characteristic length of P iis the i-th text paragraph in the quality zeroing report text; Tokenizer() is the text character splitting rule function; i is the text paragraph number, i∈1,…,n; n is the total number of text paragraphs;

[0013] S13: Determine the adjacent text paragraph P according to the characteristic length len() of the text paragraph i ,P i+1 Whether the merging conditions are met, the text paragraphs that meet the merging conditions are merged to obtain the quality zero report text block set:

[0014] B={B1,…,B j ,…,B m};

[0015] Among them, B is the quality zeroing report text block set; B j is the jth mass zeroing report text block; j is the mass zeroing report text block number; m is the total number of mass zeroing report text blocks;

[0016] S2: Extract key information feature text of aviation product quality zeroing report text and generate inspection prompts;

[0017] S21: Obtain the verification rule set r = {r1, ..., R i ,…,R n}; Parse the inspection rules of the quality zeroing report text, and for each inspection rule R i Perform component syntactic analysis to obtain the syntactic structure tree S of the test rule i for:

[0018] S i =ConGyiyuencyParGe(R i )i∈{1,2,…,n};

[0019] Among them, S i To test the rule R i The syntax tree structure of ConGyiyuencyParGe is the component syntax analysis function; R i is the i-th test rule;

[0020] S22: Identify the key information feature text of the quality zeroing report text; i Identify the noun phrase NP and the verb phrase VP, and splice them into key information feature text for the test rule R i Key information features of text I i The syntax structure S i The concatenation of noun phrases NP and verb phrases VP in the text constructs a key information feature text set I = {I1,…,Ii ,o,I n};

[0021] S23: Verify that the rule set R = {R1, R2, o, R n} Input the prompt template T in sequence, and get the test prompt as follows:

[0022]

[0023] in, To test the rule R i The inspection prompt; T is the prompt template; FillTemplaye is the inspection prompt function;

[0024] The construction test prompt set is:

[0025]

[0026] Among them, P is the test prompt set;

[0027] S3: Contextual positioning of paragraphs in the aviation product quality zeroing report text;

[0028] S31: Get the quality zeroing report text block set B in step S1 and the key information feature text set I in step S2, and use them as input sequences, and use the sorting model based on the Transformer architecture to obtain [CLS]I j [SEP]B i [SEP]; Use the similarity function Sim(B j ,I i ) to calculate each key information feature text I i With each mass zeroing report text block B j The degree of fit;

[0029] S32: Sort the text blocks based on the degree of fit obtained in step S31, and sort them from large to small according to the degree of fit, and obtain the sorted quality zero report text block set as follows:

[0030]

[0031] Among them, ranked(I i ,B) is the sorted quality zeroing report text block set; {j1,j2,…,j m} is a sequence of {1,2,…,m} that satisfies This holds true for all p=1,2,…m-1;

[0032] S33: Output the context paragraph location of the aviation product quality zero report text; the ranked quality zero report text block set ranked (I i ,B), select the first k text blocks As a candidate context paragraph, text merging is performed to locate the context text of the aviation product quality zeroing report text as follows:

[0033]

[0034] Among them, C j is the context text of the aviation product quality zeroing report text; combine is the text merging operation function; k is the text block value parameter;

[0035] S4: Use large-scale language model verification to verify the aviation product quality zeroing report text;

[0036] The inspection rule R in step S23 is respectively j Corresponding test prompt set and the context text C of the aviation product quality zeroing report text obtained in step S33 j Splice and input the large-scale language model, set the JSON format as the output format of the output inspection results; integrate and summarize the output inspection results of the quality zero report of aviation products, and output the judgment results of aviation product quality in a tabular form.

[0037] Preferably, the text character splitting rule function in step S12 is for Chinese text, and the text is split according to Chinese characters, while the English text is split according to English words.

[0038] Preferably, in step S13, the adjacent text paragraph P is determined according to the characteristic length len() of the text paragraph. i ,P i+1 Whether the merger conditions are met, specifically:

[0039]

[0040] Among them, SouldMerge is a function for determining whether the merge condition is met; True means that the result meets the merge condition; False means that the result does not meet the merge condition; L max is the maximum input text length;

[0041] For text segments that meet the merging conditions, perform the merging operation P' i =P i ∪P i+1 If SouldMerge(P i ,P i+1 )=True, use P'i To replace P i and P i+1 , then re-judge P' i and P i+2 Whether the merging conditions are met and the merging is performed.

[0042] Preferably, the component syntactic analysis in step S21 can extract noun phrases and verb phrases, as well as key information feature text from the rules.

[0043] Preferably, the key information feature text set I in step S22 is feature information extracted from the inspection rules, including time, location and fault type, with a total of n information items.

[0044] Preferably, the prompt template T in step S23 is specifically: a standard prompt file in JSON format that determines whether it meets the requirements based on the above content.

[0045] Preferably, the sorting model of the Transformer architecture in step S31 captures the relationship between the query and the candidates through a self-attention mechanism, thereby outputting the similarity.

[0046] Preferably, the output format in step S4 is in JSON format, specifically including: inspection items, inspection results and inspection basis; the inspection items are inspection rules R j ; The inspection results are divided into three categories: compliant, non-compliant and not applicable; The inspection basis is the reason for judgment.

[0047] In a second aspect of the present invention, an aviation product quality report verification system obtained by an aviation product quality report verification method based on a pre-trained language model is proposed, which includes: a zero report text segmentation module, a key information and verification prompt module, a text paragraph context positioning module, a zero report verification module and a verification result integration output module;

[0048] The zeroed report text segmentation module divides long documents into paragraphs and logic blocks, and each block adapts to the input length limit of the large-scale language model;

[0049] The key information and verification prompt module sets the key information and compliance standards for verification in the zeroing report, and classifies each item so as to judge its applicability during the verification process;

[0050] The text paragraph context location module uses a rearrangement model to reorder the divided text blocks to achieve optimal order and context coherence;

[0051] The zero report verification module uses a large-scale language model to input the report context and verification prompts to input the large-scale language model, check the compliance of the zero report, record non-conforming items and inapplicable items, and provide a basis for judgment to achieve compliance of the report;

[0052] The verification result integration output module summarizes the verification results of each zeroing report to form a unified summary table file.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] (1) The present invention uses a large-scale pre-trained language model to analyze the zero-return report on aviation product quality issues, improve the accuracy of extracting key information in the text of the report, and realize automatic verification of the report content to ensure its integrity and logical consistency.

[0055] (2) The present invention can discover common problems in aviation product quality control based on the quality zeroing report and optimize improvement plans, providing reliable data support for the improvement, optimization and standardization of aviation product quality zeroing reports.

[0056] (3) The present invention realizes zero report text segmentation, key information and verification prompts, text paragraph context positioning, zero report verification and verification result integrated output through the aviation product quality report verification system, which greatly improves the verification efficiency and accuracy of the quality zero report text. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of the aviation product quality report verification method based on the pre-trained language model of the present invention;

[0058] Figure 2 The figure is a flow chart of the aviation product quality report verification system of the present invention. DETAILED DESCRIPTION

[0059] The exemplary embodiments, features and aspects of the present invention will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0060] The embodiment of the present invention provides an aviation product quality report verification method based on a pre-trained language model, such as Figure 1 As shown, the aviation product quality zero report text is obtained, the feature length is extracted, and the merging and block operations are performed; the key information feature text of the aviation product quality zero report text is extracted to generate inspection prompts; the paragraphs of the aviation product quality zero report text are contextually located; and the aviation product quality zero report text is verified using a large-scale language model; which includes:

[0061] Step S1: Obtain the aviation product quality zeroing report text, extract the feature length, and perform merging and blocking operations.

[0062] In a preferred embodiment of the present invention, the aviation product quality zeroing report text includes five modules, namely problem location, mechanism analysis, problem reproduction, measures taken and learning from one example to another. Each module ranges from several hundred to thousands of words, and the longest may reach several thousand words. Moreover, in a specific application embodiment, for example, the zeroing report text of the technical zeroing report on the pressure swing problem of a hydraulic plunger pump contains a total of 8363 characters.

[0063] Step S11: Determine the quality zero report text and the maximum input text length; set the quality zero report text D to include n text paragraphs {P1, ...P i …,P n}, and the maximum input text length of the model is L max , limit the limit of the text block. In the embodiment, the zero report text contains a total of 190 paragraphs, set L max The value is 2000, which is greater than the length of all paragraphs.

[0064] Step S12: Use the language model to calculate each text paragraph P i The characteristic length len(P i )for:

[0065] len(P i )=len(Tokenizer(P i ));

[0066] Among them, len(P i ) is the i-th text paragraph P i The characteristic length of i is the i-th text paragraph in the quality zeroing report text; Tokenizer() is the text character splitting rule function; i is the text paragraph number, i∈1,…,n; n is the total number of text paragraphs.

[0067] The text character splitting rule function is for Chinese text, and splits the text according to Chinese characters, while for English text, it splits the text according to English words. In the embodiment, the feature lengths of all paragraphs are calculated, and the largest length is 256.

[0068] Step S13: Determine the adjacent text paragraph P according to the characteristic length len() of the text paragraph i ,P i+1 Whether the merger conditions are met, specifically:

[0069]

[0070] Among them, SouldMerge is a function for determining whether the merge condition is met; True means that the result meets the merge condition; False means that the result does not meet the merge condition; L max is the maximum input text length.

[0071] For text paragraphs that meet the merging conditions, perform the merging operation p' o =p i ∪P i+1 If SouldMerge(P i ,P i+1 )=True, use P' i To replace P i and P i+1 , then re-judge P' i and P i+2 Whether the merging conditions are met and the merging is performed, the quality zeroing report text block set is obtained as follows:

[0072] B={B1,…,B j ,…,B m};

[0073] Among them, B is the quality zeroing report text block set; B j is the jth mass zeroing report text block; j is the mass zeroing report text block number; m is the total number of mass zeroing report text blocks. In the embodiment, the zeroing report paragraph texts are merged in this way to form 5 text blocks, among which the one with the largest characteristic length is 1955.

[0074] Step S2: extract key information feature text of the aviation product quality zeroing report text and generate inspection prompts.

[0075] Step S21: Obtain the verification rule set R of the quality zeroing report text = {R1, ..., R i ,…,R n}; In the embodiment, the inspection rule set R is:

[0076]

[0077] Parse the inspection rules of the quality zeroing report text, and for each inspection rule R i Perform component syntactic analysis to extract noun phrases and verb phrases, as well as key information feature text from the rules; obtain the syntactic structure tree S of the test rules i for:

[0078] S i =ConGyiyuencyParGe(R i )i∈{1,2,…,n};

[0079] Among them, S i To test the rule R i The syntax tree structure of ConGyiyuencyParGe is the component syntax analysis function; R i is the i-th test rule.

[0080] In the embodiment, the syntax tree structure S1 is the result of the component syntax analysis of the inspection rule R1, and the result is:

[0081] (ROOT

[0082] (CP

[0083] (PP(P pair)(NP(QP(DNP(NN quality problem)(DEG's))(NN description))))

[0084] (IP

[0085] (VP

[0086] (ADVP(AD yes or no))

[0087] (VP

[0088] (VP(ADVP(AD clear))(VP(VV explanation)))

[0089] (NP

[0090] (NP(NN occurs))

[0091] (CP(IP(NP(PN's))(NP(NN time)(PU,)(NN location)(PU,)(NN timing))))))))

[0092] (PU?)))

[0093] The same operation is performed for the test rules R2 and R3.

[0094] Step S22: Identify the key information feature text of the quality zeroing report text; i Identify the noun phrase NP and the verb phrase VP, and splice them into key information feature text for the test rule R i Key information features of text I i The syntax structure S i The concatenation of noun phrases NP and verb phrases VP in the text constructs a key information feature text set I = {I1,…,I i ,…,I n}, there are n information items in total.

[0095] In the embodiment, the key information item I1 obtained by checking the rule R1 and the syntax tree structure S1 is "quality problem description, time, location, timing", and similarly, I2 and I3 can be obtained as "problem product, structure, function, performance" and "fault tree analysis" respectively.

[0096] Step S23: Verify that the rule set R = {R1, R2, ..., R n} Input the prompt template T in sequence, and get the test prompt as follows:

[0097]

[0098] in, To test the rule R i T is the prompt template, which determines whether it meets the requirements of the standard prompt file in JSON format according to the content above; FillTemplaye is the prompt test function.

[0099] The construction test prompt set is:

[0100]

[0101] Among them, P is the test prompt set.

[0102] Step S3: Contextually locate the paragraphs of the aviation product quality zeroing report text.

[0103] Step S31: Obtain the quality zeroing report text block set B in step S1 and the key information feature text set I in step S2, and use them as input sequences, and use the sorting model based on the Transformer architecture to obtain [CLS]I j [SEP]B i [SEP]; The Transformer architecture ranking model captures the relationship between the query and the candidate through the self-attention mechanism, thereby outputting the similarity. Using the similarity function Sim(B j ,I i ) to calculate each key information feature text I i With each mass zeroing report text block B j degree of fit.

[0104] Step S32: Sort the text blocks based on the degree of fit obtained in step S31, and sort them from large to small according to the degree of fit, and obtain the sorted quality zero report text block set as follows:

[0105]

[0106] Among them, ranked(I i ,B) is the sorted quality zeroing report text block set; {j1,j2,…,jm} is a sequence of {1,2,…,m} that satisfies This holds true for all p=1,2,…m-1.

[0107] Step S33: Output the context paragraph location of the aviation product quality zero report text; the ranked quality zero report text block set ranked (I i ,B), select the first k text blocks As a candidate context paragraph, text merging is performed to locate the context text of the aviation product quality zeroing report text as follows:

[0108]

[0109] Among them, C j is the context text of the aviation product quality zeroing report text; combine is the text merging operation function; k is the text block value parameter, and the default value is 1 or 2.

[0110] Step S4: Use a large-scale language model to verify the aviation product quality zeroing report text.

[0111] The inspection rule R in step S23 is respectively j Corresponding test prompt set and the context text C of the aviation product quality zeroing report text obtained in step S33 j Splice and input the large-scale language model, set the JSON format as the output format of the output test results; the output format JSON format specifically includes: test items, test results and test basis; the test item is the test rule R j ; The inspection results are compliant, noncompliant and not applicable; the inspection basis is the reason for judgment. In the prompt, it is stipulated that the large model is output in the following format according to the JSON format:

[0112] {"Check item":"R j ", / / Verification rule R j

[0113] "Verification result": "Compliant" | "Not Compliant" | "Not Applicable", / / Verification result

[0114] "Basis":"Reason for judgment" / / Basis for judgment}

[0115] Set R1, R2, ..., R n and its corresponding The large-scale language model is input in this way in sequence. In the embodiment, the model output result is finally obtained in the following format:

[0116] [{"Verification Item":"Whether the description of the quality problem clearly states the time, place, and timing of occurrence","Verification Result":"Not compliant","Basis":"The zeroing report does not clearly state the time, place, timing, phenomenon (consequences), and environmental conditions of the occurrence. The report mainly focuses on the mechanism analysis, fault tree analysis, and corrective measures of the problem, but does not describe the specific time and location of occurrence in detail."},

[0117] {"Check item":"Does the structure, working principle, function, performance, etc. of the problem product have been briefly introduced?","Check result":"Not in compliance","Basis":"The zeroing report mainly focuses on fault analysis, location and mechanism analysis, and does not provide detailed information on the structure, working principle, function, performance, etc. of the problem product."},{"Check item":"Has a fault tree analysis been established?","Check result":Compliant","Basis":"Based on the content of the zeroing report provided, a fault tree analysis has indeed been established. The report clearly mentions the fault tree with fatigue cracks in the rotor assembly as the top event, and lists each branch and factor in detail."}]

[0118] Integrate and summarize the output inspection results of the zero-quality report of aviation products, and output the judgment results of aviation product quality in a table. Integrate and summarize the verification results of the zero-quality report so that management can effectively analyze the compliance of each verification item and find common problems and improvement points. The output of this step is a unified summary table showing the verification results of each report, as shown in Table 1, including the compliance of each verification item, whether it meets the requirements, and its judgment basis.

[0119] Table 1 Output of the judgment results of aviation product quality

[0120]

[0121] Then parse the JSON format results output in S4 and fill in the corresponding items in the results into the table. For different reports, fill in the report number, verification items, verification results and basis in this way. Finally, save and output the table in Excel format to form the final output result of the system.

[0122] The second aspect of the present invention proposes an aviation product quality report verification system obtained by an aviation product quality report verification method based on a pre-trained language model. Relying on a large language model, the system realizes the automatic verification, recognition and information extraction of the quality zero report, greatly improving the verification efficiency and accuracy of the quality zero report text. The system includes: a zero report text segmentation module, a key information and verification prompt module, a text paragraph context positioning module, a zero report verification module and a verification result integration output module; the overall process of the system is as follows Figure 2 shown.

[0123] The text segmentation module of Zero Report divides long documents into paragraphs and logical chunks, ensuring that each chunk is logically complete and fits the input length limit of large-scale language models.

[0124] The key information and verification prompt module sets the key information and compliance standards that must be verified in the zeroing report, and classifies each item to determine its applicability during the verification process.

[0125] The text paragraph context localization module uses a reordering model to reorder the segmented text blocks to ensure the optimal order and contextual coherence of task processing.

[0126] Zero report verification The verification module uses a large-scale language model to input the report context and verification prompts into the large-scale language model to check the compliance of the zero report, record non-conforming items and inapplicable items, and provide a basis for judgment to ensure the integrity and compliance of the report.

[0127] The verification result integration output module summarizes the verification results of each zeroing report and designs a unified summary table to facilitate management to analyze the compliance status of each verification project and identify common problems and improvement points.

[0128] The beneficial effects of the present invention are as follows: the present invention provides an aviation product quality report verification method based on a pre-trained language model, uses a large-scale pre-trained language model to analyze the aviation product quality problem zeroing report, improves the accuracy of extracting key information in the text of the report, and realizes automatic verification of the report content to ensure its integrity and logical consistency; discovers common problems in aviation product quality control based on the quality zeroing report and optimizes improvement plans, providing reliable data support for the improvement, optimization and standardization of aviation product quality zeroing reports. In the embodiment, the practicality of the present invention is verified by the aviation product quality report verification system, which greatly improves the verification efficiency and accuracy of the quality zeroing report text.

[0129] The embodiments described above are only descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for verifying aviation product quality reports based on a pre-trained language model, characterized in that: It includes: S1: Obtain the aviation product quality zeroing report text, extract the feature length, and perform merging and block operations; S11: Determine the mass zeroing report text and the maximum input text length; set the mass zeroing report text D to contain n text paragraphs {P1,…P i …,P n }, and the maximum input text length of the model is L max ,limit the limit of text blocks; S12: Use the language model to calculate each text paragraph P i The characteristic length len(P i )for: len(P i )=len(Tokenizer(P i )) Among them, len(P i ) is the i-th text paragraph P i The characteristic length of i is the i-th text paragraph in the quality zeroing report text; Tokenizer() is the text character splitting rule function; i is the text paragraph number, i∈1,…,n; n is the total number of text paragraphs; S13: Determine the adjacent text paragraph P according to the characteristic length len() of the text paragraph i ,P i+1 Whether the merging conditions are met, the text paragraphs that meet the merging conditions are merged to obtain the quality zero report text block set: B={B1,…,B j ,…,B m } Among them, B is the quality zeroing report text block set; B j is the jth mass zeroing report text block; j is the mass zeroing report text block number; m is the total number of mass zeroing report text blocks; S2: Extract key information feature text of aviation product quality zeroing report text and generate inspection prompts; S21: Obtain the inspection rule set R = {R1, ..., R i ,…,R n }; Parse the inspection rules of the quality zeroing report text, and for each inspection rule R i Perform component syntactic analysis to obtain the syntactic structure tree S of the test rule i for: S i =ConstituencyParse(R i ) i∈{1,2,…,n} Among them, S i To test the rule R i The syntactic tree structure of R i is the i-th test rule; S22: Identify the key information feature text of the quality zeroing report text; i Identify the noun phrase NP and the verb phrase VP, and splice them into key information feature text for the test rule R i Key information features of text I i The syntax structure S i The concatenation of noun phrases NP and verb phrases VP in the text constructs a key information feature text set I = {I1,…,I i ,…,I n }; S23: Verify that the rule set R = {R1, R2, ..., R n } Input the prompt template T in sequence, and get the test prompt as follows: in, To test the rule R i Test prompt; T is the prompt template; FillTemplate is the test prompt function; The construction test prompt set is: Among them, P is the test prompt set; S3: Contextual positioning of paragraphs in the aviation product quality zeroing report text; S31: Get the quality zeroing report text block set B in step S1 and the key information feature text set I in step S2, and use them as input sequences, and use the sorting model based on the Transformer architecture to obtain [CLS]I j [SEP]B i [SEP]; Use the similarity function Sim(B j ,I i ) to calculate each key information feature text I i With each mass zeroing report text block B j The degree of fit; S32: Sort the text blocks based on the degree of fit obtained in step S31, and sort them from large to small according to the degree of fit, and obtain the sorted quality zero report text block set as follows: Among them, ranked(I i ,B) is the sorted quality zeroing report text block set; {j1,j2,…,j m } is a sequence of {1,2,…,m} that satisfies This holds true for all p=1,2,…m-1; S33: Output the context paragraph location of the aviation product quality zero report text; the ranked quality zero report text block set ranked (I i ,B), select the first k text blocks As a candidate context paragraph, text merging is performed to locate the context text of the aviation product quality zeroing report text as follows: Among them, C j is the context text of the aviation product quality zeroing report text; combine is the text merging operation function; k is the text block value parameter; S4: Use large-scale language model verification to verify the aviation product quality zeroing report text; The inspection rule R in step S23 is respectively j Corresponding test prompt set and the context text C of the aviation product quality zeroing report text obtained in step S33 j Splice and input the large-scale language model, set the JSON format as the output format of the output inspection results; integrate and summarize the output inspection results of the quality zero report of aviation products, and output the judgment results of aviation product quality in a tabular form.

2. The aviation product quality report verification method based on the pre-trained language model according to claim 1 is characterized in that: In step S12, the text character splitting rule function splits the Chinese text according to Chinese characters, and splits the English text according to English words.

3. The aviation product quality report verification method based on the pre-trained language model according to claim 1 is characterized in that: In step S13, the adjacent text paragraph P is determined according to the characteristic length len() of the text paragraph. i ,P i+1 Whether the merger conditions are met, specifically: Among them, SouldMerge is a function for determining whether the merge condition is met; True means that the result meets the merge condition; False means that the result does not meet the merge condition; L max is the maximum input text length; For text segments that meet the merging conditions, perform the merging operation P' i =P i ∪P i+1 If SouldMerge(P i ,P i+1 )=True, use P' i To replace P i and P i+1 , then re-judge P' i and P i+2 Whether the merging conditions are met and the merging is performed.

4. The aviation product quality report verification method based on the pre-trained language model according to claim 1 is characterized in that: In step S21, through component syntactic analysis, noun phrases and verb phrases, as well as key information feature texts can be extracted from the rules.

5. The aviation product quality report verification method based on the pre-trained language model according to claim 1 is characterized in that: Step S22: The key information feature text set I is feature information extracted from the inspection rules, including time, location and fault type, with a total of n information items.

6. The aviation product quality report verification method based on the pre-trained language model according to claim 1 is characterized in that: The prompt template T in step S23 is specifically: a standard prompt file in JSON format that determines whether it meets the requirements based on the above content.

7. The aviation product quality report verification method based on the pre-trained language model according to claim 1 is characterized in that: In step S31, the sorting model of the Transformer architecture captures the relationship between the query and the candidate through the self-attention mechanism, thereby outputting the similarity.

8. The aviation product quality report verification method based on the pre-trained language model according to claim 1 is characterized in that: The output format in step S4 is JSON format, which specifically includes: inspection items, inspection results and inspection basis; the inspection items are inspection rules R j ; The inspection results are divided into three categories: compliant, non-compliant and not applicable; The inspection basis is the reason for judgment.

9. An aviation product quality report verification system obtained by an aviation product quality report verification method based on a pre-trained language model, characterized in that: It includes: Zero report text segmentation module, key information and verification prompt module, text paragraph context positioning module, zero report verification and validation module and verification result integration and output module; The zeroed report text segmentation module divides long documents into paragraphs and logic blocks, and each block adapts to the input length limit of the large-scale language model; The key information and verification prompt module sets the key information and compliance standards for verification in the zeroing report, and classifies each item so as to judge its applicability during the verification process; The text paragraph context location module uses a rearrangement model to reorder the divided text blocks to achieve optimal order and context coherence; The zero report verification module uses a large-scale language model to input the report context and verification prompts to input the large-scale language model, check the compliance of the zero report, record non-conforming items and inapplicable items, and provide a basis for judgment to achieve compliance of the report; The verification result integration output module summarizes the verification results of each zeroing report to form a unified summary table file.

Images