Aerospace equipment fault tree generation method based on large model

Abstract

The invention relates to an aerospace equipment fault tree generation method based on a large model, and belongs to the technical field of aerospace reliability. By analyzing multi-source heterogeneous information, a current process step is quickly positioned, and on this basis, a user is assisted in obtaining related information such as operation detailed rules, disposal plans, design mechanisms, historical faults and matters needing attention, so that information pushing and knowledge accompanying are realized. The technology supports on-line monitoring, trend prediction and abnormity early warning of real-time data, and fault attribution and auxiliary troubleshooting are realized by acquiring troubleshooting knowledge and related data and assisting fault phenomenon description input by a user when abnormity occurs; the problems that the multi-source heterogeneous data is difficult to effectively integrate and utilize, the text data search efficiency is low, the quality is poor and the troubleshooting knowledge potential tapping is insufficient are solved.

Description

Technical Field

[0001] The invention belongs to the technical field of aerospace reliability and relates to a method for generating a fault tree of aerospace equipment based on a large model. Background Art

[0002] Aerospace systems have a trend of becoming more complex in structure, more complicated in testing and launching processes, and more severe in environment. Therefore, intelligent troubleshooting technology can significantly improve testing and launching efficiency, and puts forward higher requirements for the intelligence and reliability of software. When a system fails, it is necessary to perform troubleshooting efficiently and accurately, list possible causes of the failure, locate the location of the failure, and the order and method of troubleshooting.

[0003] However, there is currently no relevant method, and it is impossible to achieve intelligent troubleshooting based on the current status. Summary of the invention

[0004] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art and propose a method for generating a fault tree for aerospace equipment based on a large model, which solves the problems of difficulty in effectively integrating and utilizing multi-source heterogeneous data, low efficiency and poor quality of text data search, and insufficient potential of troubleshooting knowledge.

[0005] The solution of the present invention is:

[0006] A method for generating a fault tree of aerospace equipment based on a large model, comprising:

[0007] Collect text and image data related to processes and troubleshooting; perform text preprocessing and image preprocessing to form preprocessed data;

[0008] Annotate the text and images of the preprocessed data that describe the same content or related entities;

[0009] Select a pre-trained multimodal model and fine-tune the model using aerospace corpus;

[0010] The description information input by the user, the current status monitoring information, and the real-time monitoring image information are processed using a multimodal model after aerospace corpus alignment, and the text and image data are converted into a unified vector format for representation;

[0011] Use Euclidean distance calculation to measure the similarity of vectors and achieve state positioning;

[0012] Build a troubleshooting knowledge graph;

[0013] Pre-build an aerospace vector knowledge base and store candidate vectors in the vector knowledge base; use the candidate vectors and state description vectors as inputs, use the re-ranking model to calculate the vector similarity, and obtain the associated knowledge;

[0014] Based on the relevant knowledge obtained, information push and knowledge accompaniment based on the status are realized to provide information support for users;

[0015] Set up coding rules;

[0016] Write prompt words;

[0017] Check the output results of the large model and regenerate them if there are any problems;

[0018] The verified large model generation results are pushed to the user to realize fault tree generation and auxiliary troubleshooting.

[0019] In the above-mentioned method for generating a fault tree for aerospace equipment based on a large model, the text preprocessing includes space removal and paragraph division; the image preprocessing includes image noise removal and image alignment.

[0020] In the above-mentioned method for generating a fault tree for aerospace equipment based on a large model, the annotated content includes key information and association relationships.

[0021] In the above-mentioned method for generating aerospace equipment fault tree based on a large model, the model is fine-tuned through aerospace corpus to achieve corpus alignment between texts, between images and texts, and between images and images.

[0022] In the above-mentioned method for generating a fault tree for aerospace equipment based on a large model, the method for constructing the fault troubleshooting knowledge graph is as follows:

[0023] Based on the aerospace corpus system, related plans, details, process steps, and historical data information are associated through vector relevance to construct a troubleshooting knowledge graph.

[0024] In the above-mentioned method for generating aerospace equipment fault tree based on a large model, the method for constructing the aerospace vector knowledge base is as follows:

[0025] S11. Acquire the knowledge system in the specified field;

[0026] S12, preprocessing the knowledge in the specified field;

[0027] S13, inputting the knowledge of the specified field into a pre-trained multimodal model to obtain a vector representation;

[0028] S14, knowledge clustering through vector similarity;

[0029] S15. Build a vector knowledge base based on the clustering results.

[0030] In the above-mentioned method for generating aerospace equipment fault tree based on a large model, the calculation method of vector similarity is:

[0031] S21. Obtain multiple pieces of relevant information, and manually mark the relevance of the information;

[0032] S22, selecting two pieces of information synergistically as inputs of a reordering model, and training the reordering model;

[0033] S23, training the similarity prediction model to be trained according to the similarity marked by the user to reduce the similarity training target output by the reordering model;

[0034] S24. Through the pre-trained re-ranking model, the text information input by the user and other information in the collaborative troubleshooting knowledge base are taken as input to achieve similarity generation.

[0035] In the above-mentioned method for generating a fault tree for aerospace equipment based on a large model, the fault tree in image form is annotated into text form by designing a fault tree coding rule, and the fault tree output by the large model complies with the same coding rule.

[0036] In the above-mentioned method for generating a fault tree for aerospace equipment based on a large model, the method for writing prompt words is: using the acquired relevant knowledge and fault tree coding rules as prompt words, combined with the fault phenomenon description information input by the user as the large model input, to obtain the analysis result generated by the large model.

[0037] In the above-mentioned method for generating a fault tree for aerospace equipment based on a large model, the content of checking the output result of the large model includes whether the fault tree coding rules are compliant and whether the content generated by the large model conflicts with the actual situation.

[0038] The beneficial effects of the present invention compared with the prior art are:

[0039] (1) The large-model-based aerospace equipment fault tree generation method provided by the present invention can realize rapid status positioning based on monitoring data, user description information and other data;

[0040] (2) The present invention combines relevant knowledge to realize information push and knowledge accompaniment, and generates a fault tree to complete fault attribution and assist in troubleshooting when an abnormality occurs;

[0041] (3) The present invention supports online monitoring, trend prediction and abnormal warning of real-time data. When an abnormality occurs, it can achieve fault attribution and auxiliary troubleshooting by acquiring troubleshooting knowledge and related data, supplemented by the fault phenomenon description input by the user. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The present invention provides a flow chart for generating a fault tree for aerospace equipment based on a large model. DETAILED DESCRIPTION

[0043] The present invention will be further described below in conjunction with the embodiments.

[0044] The present invention provides a method for generating a fault tree for aerospace equipment based on a large model. By analyzing multi-source heterogeneous information, the current process steps can be quickly located, and on this basis, the user can be assisted in obtaining relevant information such as operation details, disposal plans, design mechanisms, historical faults, precautions, etc., to achieve information push and knowledge accompaniment. This technology supports online monitoring, trend prediction and abnormal warning of real-time data. When an abnormality occurs, it can achieve fault attribution and auxiliary troubleshooting by obtaining troubleshooting knowledge and related data, supplemented by the description of the fault phenomenon input by the user.

[0045] The present invention provides a method for generating a fault tree of aerospace equipment based on a large model, such as Figure 1 As shown, there are 3 key steps:

[0046] (1) Design heterogeneous information association and fusion technology based on multimodal models to realize current status identification by combining the status description in the operation rules, disposal plan, and zeroing report with the actual monitoring data and image information.

[0047] Step 1: Data collection: collect a large amount of text and image data related to processes and troubleshooting, and perform text preprocessing operations such as space removal and paragraph division, and image preprocessing operations such as image noise removal and image alignment;

[0048] Step 2: Data annotation: select some of the collected data and annotate the text and images that describe the same content or related entities, including key information and associations;

[0049] Step 3: Corpus alignment: select a pre-trained multimodal model and fine-tune the model using the aerospace corpus to achieve corpus alignment between "text-text", "image-text", and "image-image";

[0050] Step 4: State vectorization: Use the multimodal model after aerospace corpus alignment to process the user input description information, current state monitoring information, real-time monitoring image information and other data, and convert text, image and other data into a unified vector format for representation;

[0051] Step 5: State association, use the Euclidean distance calculation method to measure the similarity of vectors and achieve state positioning.

[0052] (2) Design a retrieval enhancement generation method for multimodal semantic matching to achieve state-based information push and knowledge accompaniment, which includes the following steps:

[0053] Step 1: Knowledge construction. Based on the aerospace corpus system obtained in step (1), the related plans, details, process steps, historical data and other information are associated through vector relevance to build a troubleshooting knowledge graph.

[0054] Step 2: Information acquisition: The candidate vectors stored in the pre-built vector knowledge base are used together with the state description vector obtained in step (1) as input, and the vector similarity is calculated using the re-ranking model, and the associated information is obtained on this basis;

[0055] Step 3: Knowledge push. Based on the relevant knowledge obtained in step 2, information push and knowledge accompaniment based on the status are implemented to provide information support for users.

[0056] The method described in step 2 is characterized by: 1. Pre-constructing an aerospace knowledge vector library, specifically including: ① acquiring a knowledge system in a specified field; ② pre-processing the knowledge in the specified field; ③ inputting the knowledge in the specified field into a pre-trained multimodal model to obtain a vector representation; ④ realizing knowledge clustering through vector similarity; ⑤ constructing an aerospace knowledge system based on the clustering results.

[0057] The method described in step 2 is characterized by: 2. Pre-training a re-ranking model to achieve similarity measurement, specifically including: ① obtaining multiple relevant information and manually marking the relevance of the information; ② selecting two pieces of information to be coordinated as inputs of the re-ranking model and training the re-ranking model; ③ training the similarity prediction model to be trained based on the similarity marked by the user to reduce the similarity training target output by the re-ranking model. ④ Through the pre-trained re-ranking model, the text information input by the user and other information in the collaborative troubleshooting knowledge base are used as inputs to achieve similarity generation.

[0058] (3) Design a fault tree generation technology based on a large model. Utilize the relevant information generated in step (2) and use the large model technology to tap the potential troubleshooting knowledge therein. Analyze the possible causes of the fault based on the fault phenomenon description input by the user and generate a fault tree to assist the user in troubleshooting the possible location of the fault and improve the troubleshooting efficiency. The steps are as follows:

[0059] Step 1: Coding rule setting. By designing the fault tree coding rules, the fault tree in image form is annotated into text form. The fault tree output by the large model should comply with the same coding rules.

[0060] Step 2: Prompt word writing: the relevant knowledge obtained in step (2) and the fault tree coding rules are used as prompt words, combined with the fault phenomenon description information input by the user as the input of the large model, to obtain the analysis results generated by the large model;

[0061] Step 3: Result verification: Verify the output results of the large model, including whether the fault tree coding rules are compliant, whether the generated content of the large model conflicts with the actual situation, etc. If there are any problems, regenerate;

[0062] Step 4: Information push: push the verified large model generation results to the user to achieve fault tree generation and auxiliary troubleshooting.

[0063] The present invention also designs the following contents:

[0064] Retrieval Enhanced Generative Method

[0065] Step 1: Obtain the text to be processed and determine the feature vector of the text to be processed.

[0066] In this example, the text to be processed may be information input by the user, and the questions or problems described in the text to be processed may be related to a specified field. The specified field may be a highly professional technical field, such as legal consulting, medical medicine, financial analysis, engineering design and other knowledge-intensive fields. There is a lot of professional knowledge in the specified field, and the knowledge is relatively complex. The length of the text to be processed can be determined according to the specific application scenario.

[0067] Further, a pre-trained word vector embedding model can be used to determine the feature vector of the text to be processed. Specifically, the word vector embedding model used in this example is obtained by training with the general corpus in the general corpus. The text to be processed is input into the pre-trained word vector embedding model to obtain the feature vector of the text to be processed output by the model. The word vector embedding model can be a structure such as Word2Vec, GloVe, FastText, etc., wherein the word vector embedding model of the Word2Vec structure can adopt a continuous bag of words model (CBOW) or a skip-gram model (Skip-gram). This example does not specifically limit the model structure of the word vector embedding model. In addition, the training sample of the word vector embedding model can be derived from a general corpus. The general corpus is subjected to pre-processing operations such as word segmentation and stop word removal, and converted into a form recognizable by the word vector embedding model. After that, the general corpus after processing is used as a training sample to train the word vector embedding model to obtain a word vector embedding model after training.

[0068] It should be noted that in this example, the word vector embedding model is not optimized and adjusted to adapt to the specified field. The word vector embedding models used to determine the feature vector or word vector of the text in step 1 and subsequent steps are all trained based on the above method, which will not be repeated later.

[0069] Step 2: Input the candidate vectors corresponding to the candidate texts stored in the pre-built specified vector database and the feature vector of the text to be processed as inputs to the pre-trained similarity prediction model to obtain the similarity between the candidate vectors and the feature vector of the text to be processed. The pre-trained similarity prediction model is trained using text pairs in the specified field as training samples and the similarity of the text pairs in the specified field as annotations.

[0070] Specifically, the difference between the retrieval enhancement generation method based on vector similarity matching optimization provided in this example and the traditional retrieval enhancement generation method is that the traditional method uses similarity algorithms to determine vector similarity, such as Euclidean distance, dot product, cosine similarity, etc. In this example, the vector similarity is determined by using a similarity prediction model that uses text pairs in a specified field as training samples and the similarity of the text pairs in the specified field as annotation training.

[0071] The similarity prediction model trained with text pairs in a specified field as training samples can accurately determine the vector similarity corresponding to the text in the specified field, so that more matching professional knowledge in the specified field can be retrieved for the text to be processed.

[0072] Specifically, a pre-built designated vector database stores a plurality of candidate texts and candidate vectors corresponding to each candidate text. For each candidate vector, the candidate vector and the feature vector of the text to be processed are used as inputs to a pre-trained similarity prediction model to obtain the similarity between the candidate vector output by the model and the feature vector of the text to be processed. Thus, based on the similarity prediction model trained in the designated field, the similarity between each candidate vector in the vector database and the feature vector of the text to be processed is obtained.

[0073] Step 3: Determine a reference text associated with the text to be processed from the candidate texts according to the similarity between the candidate vector and the feature vector of the text to be processed.

[0074] Furthermore, based on the similarity between each candidate vector and the feature vector of the text to be processed, a reference text associated with the text to be processed is selected from each candidate text. Specifically, a preset similarity threshold is obtained, and candidate vectors with a similarity greater than the preset similarity threshold are screened out, and based on the correspondence between the candidate vectors and the candidate texts stored in a specified vector database, the candidate text corresponding to the screened candidate vector is determined as the reference text of the text to be processed. Alternatively, the candidate vectors can be arranged in order from large to small according to the similarity, and several candidate vectors with large similarity are used as the screened candidate vectors. After that, based on the correspondence between the candidate vectors and the candidate texts stored in the specified vector database, the candidate text corresponding to the screened candidate vector is determined as the reference text of the text to be processed.

[0075] The number of reference texts of the to-be-processed text retrieved based on the above scheme and the text length of each reference text are not limited in this example.

[0076] Step 4: Determine a prompt text based on the text to be processed and the reference text, and input the prompt text into a pre-trained natural language model to obtain a target text corresponding to the text to be processed.

[0077] Specifically, the pre-trained natural language model is obtained based on a general corpus. The pre-trained natural language model has good in-context learning (ICL) capability. Therefore, the retrieved reference text can be used as context information and inserted into the corresponding position of the preset prompt template, and the text to be processed can also be inserted into the corresponding position of the prompt template to obtain the prompt text.

[0078] The prompt text containing the text to be processed and the reference text is used as the input of the pre-trained natural language model, and the text generated by the pre-trained natural language model is the target text corresponding to the text to be processed.

[0079] Large Language Model Fine-tuning Method

[0080] Step 1: Determine the target pre-trained language model to be fine-tuned.

[0081] In this embodiment, the target pre-trained language model to be fine-tuned is first determined. It should be noted that the pre-trained language model may include but is not limited to GPT, BERT (Bidirectional Encoder from Transformer), T5 (Text to Text Transformer), BART (Bidirectional and Autoregressive Transformer), etc., that is, the pre-trained language model can be a language model of any type and scale, including but not limited to language models based on Transformer or other architectures, based on English or other languages, based on text or other media.

[0082] In a specific implementation method, an externally imported custom large language model is obtained to obtain a target pre-trained language model to be fine-tuned. It is understandable that if a user needs to use a custom large language model for fine-tuning, the user-imported custom large language model can be obtained through a corresponding interface, wherein importing a pre-trained model can significantly reduce training time and resource consumption.

[0083] Step 2: Select a corresponding target fine-tuning algorithm from a pre-built fine-tuning algorithm library, and set corresponding target hyperparameters for the target pre-trained language model to be fine-tuned and the target fine-tuning algorithm.

[0084] In this embodiment, the corresponding fine-tuning algorithm and its algorithm parameters are loaded or downloaded from a pre-built fine-tuning algorithm library. For example, the fine-tuning algorithms that can be selected may include, but are not limited to, full fine-tuning, Adapter Tuning, P-Tuning (i.e., a fine-tuning method for improving pre-trained language models), LoRA (i.e., a low-resource fine-tuning large model method), QLoRA (i.e., an efficient large model fine-tuning method), and other mainstream fine-tuning algorithms. Specifically, according to the second selection instruction issued by the current user on the visual interactive interface by clicking or inputting, the corresponding target fine-tuning algorithm is selected from the pre-built fine-tuning algorithm library. In other words, according to the user's input or selection, the corresponding fine-tuning algorithm and its parameters are loaded or downloaded from the pre-built fine-tuning algorithm library. In this embodiment, various classic and custom fine-tuning algorithms are supported, such as full fine-tuning, adapter, etc., thereby improving the flexibility of fine-tuning.

[0085] It should be pointed out that Adapter Tuning is a fine-tuning method for improving pre-trained language models. In this method, the pre-trained parameters of the model remain unchanged, and small adapters are inserted between the layers of the model. Then only the parameters of these adapters are fine-tuned, rather than the parameters of the entire model. P-Tuning is a fine-tuning method for improving pre-trained language models. Unlike traditional fine-tuning methods, P-Tuning introduces a task called "P-Task", such as a text classification task, and in the fine-tuning process, by introducing a weight parameter, the information of the task is fused with the pre-trained parameters of the model. LoRA fine-tuning is a low-resource method for fine-tuning large models. Based on the inherent low-rank characteristics of large models, LoRA adds a bypass matrix to simulate full-parameter fine-tuning. It is currently one of the most common and effective fine-tuning methods, and can be effectively combined with other parameter-efficient fine-tuning methods. QLoRA is an efficient large model fine-tuning method that can reduce memory usage to a level sufficient to fine-tune a 65 billion parameter model on a single 48GB GPU while maintaining full 16-bit fine-tuning task performance. QLoRA backpropagates gradients to a low-rank adapter (LoRA) through a frozen 4-bit quantized pre-trained language model.

[0086] The present invention sets corresponding target hyperparameters for the target pre-trained language model to be fine-tuned and the target fine-tuning algorithm. It is understandable that various parameters in the large model fine-tuning process are set and adjusted according to task requirements. Specifically, according to the setting instructions issued by the current user through clicking or input on the visual interactive interface, corresponding target hyperparameters are set for the target pre-trained language model to be fine-tuned and the target fine-tuning algorithm. That is to say, according to the user's input or selection, appropriate hyperparameters are set for the PLM and the fine-tuning algorithm, and various parameters such as hyperparameters and model parameters in the fine-tuning process are configured. The hyperparameters may include but are not limited to learning rate, batch size, number of iterations, regularization coefficient, etc.

[0087] Step 3: Get a pre-built training dataset corresponding to the downstream specific task.

[0088] In this embodiment, a training data set corresponding to a specific downstream task is obtained to improve the performance of the model for the specific downstream task through fine-tuning.

[0089] Step 4: According to the target fine-tuning algorithm and the target hyperparameters, and using the training data set, the target pre-trained language model to be fine-tuned is cyclically trained to obtain a corresponding target large language model.

[0090] The large-model-based aerospace equipment fault tree generation method provided by the present invention can realize rapid status positioning according to monitoring data, user description information and other data.

[0091] The present invention combines relevant knowledge to realize information push and knowledge accompaniment, and generates a fault tree to complete fault attribution and auxiliary troubleshooting when an abnormality occurs.

[0092] The present invention supports online monitoring, trend prediction and abnormal warning of real-time data. When an abnormality occurs, it can achieve fault attribution and auxiliary troubleshooting by acquiring troubleshooting knowledge and related data, supplemented by the fault phenomenon description input by the user.

[0093] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications to the technical solution of the present invention by using the methods and technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for generating a fault tree for aerospace equipment based on a large model, characterized in that: include: Collect text and image data related to processes and troubleshooting; And perform text preprocessing and image preprocessing; Forming preprocessed data; Annotate the text and images of the preprocessed data that describe the same content or related entities; Select a pre-trained multimodal model and fine-tune the model using aerospace corpus; The description information input by the user, the current status monitoring information, and the real-time monitoring image information are processed using a multimodal model after aerospace corpus alignment, and the text and image data are converted into a unified vector format for representation; Use Euclidean distance calculation to measure the similarity of vectors and achieve state positioning; Build a troubleshooting knowledge graph; Pre-build an aerospace vector knowledge base and store candidate vectors in the vector knowledge base; use the candidate vectors and state description vectors as inputs, use the re-ranking model to calculate the vector similarity, and obtain the associated knowledge; Based on the relevant knowledge obtained, information push and knowledge accompaniment based on the status are realized to provide information support for users; Set up coding rules; Write prompt words; Verify the output of the large model; If there is a problem, regenerate; The verified large model generation results are pushed to the user to realize fault tree generation and auxiliary troubleshooting.

2. The method for generating a fault tree of aerospace equipment based on a large model according to claim 1, characterized in that: The text preprocessing includes space removal and paragraph division; the image preprocessing includes image noise removal and image alignment.

3. The method for generating a fault tree for aerospace equipment based on a large model according to claim 1, characterized in that: The annotated content includes key information and association relationships.

4. The method for generating a fault tree of aerospace equipment based on a large model according to claim 1, characterized in that: The model is fine-tuned using aerospace corpus to achieve corpus alignment between texts, images, and texts, and between images and images.

5. The method for generating a fault tree of aerospace equipment based on a large model according to claim 1, characterized in that: The construction method of the troubleshooting knowledge graph is as follows: Based on the aerospace corpus system, related plans, details, process steps, and historical data information are associated through vector relevance to construct a troubleshooting knowledge graph.

6. The method for generating a fault tree of aerospace equipment based on a large model according to claim 1, characterized in that: The construction method of aerospace vector knowledge base is: S11. Acquire the knowledge system in the specified field; S12, preprocessing the knowledge in the specified field; S13, inputting the knowledge of the specified field into a pre-trained multimodal model to obtain a vector representation; S14, knowledge clustering through vector similarity; S15. Build a vector knowledge base based on the clustering results.

7. The method for generating a fault tree of aerospace equipment based on a large model according to claim 1, characterized in that: The calculation method of vector similarity is: S21. Obtain multiple pieces of relevant information, and manually mark the relevance of the information; S22, selecting two pieces of information synergistically as inputs of a reordering model, and training the reordering model; S23, training the similarity prediction model to be trained according to the similarity marked by the user to reduce the similarity training target output by the reordering model; S24. Through the pre-trained re-ranking model, the text information input by the user and other information in the collaborative troubleshooting knowledge base are taken as input to achieve similarity generation.

8. The method for generating a fault tree of aerospace equipment based on a large model according to claim 1, characterized in that: By designing fault tree coding rules, the fault tree in image form is annotated into text form, and the fault tree output by the large model conforms to the same coding rules.

9. The method for generating a fault tree of aerospace equipment based on a large model according to claim 8, characterized in that: The method for writing prompt words is: using the acquired relevant knowledge and fault tree coding rules as prompt words, combined with the fault phenomenon description information input by the user as the input of the large model, to obtain the analysis results generated by the large model.

10. The method for generating a fault tree of aerospace equipment based on a large model according to claim 9, characterized in that: The inspection of the output results of the large model includes whether the fault tree coding rules are compliant and whether the content generated by the large model conflicts with the actual situation.

Images