A UAV equipment training method based on large model and incremental learning

Abstract

The present invention discloses a method for training unmanned aerial vehicle equipment based on a large model and incremental learning, which belongs to the technical field of big data intelligence, and includes an unmanned aerial vehicle equipment knowledge base generation module, an unmanned aerial vehicle equipment training exercise recommendation module and an unmanned aerial vehicle equipment knowledge update module; the unmanned aerial vehicle equipment knowledge base generation module is constructed by a large language model, and a knowledge base is generated accordingly; the unmanned aerial vehicle equipment training exercise recommendation module relies on the knowledge base, uses an exercise generator to generate exercises, and the exercises are combined into an exercise bank, and the exercise bank is used to simulate the unmanned aerial vehicle equipment operation learning curve of the trainees and screen the unmanned aerial vehicle equipment training exercises; the unmanned aerial vehicle equipment knowledge update module adds incremental learning to realize the generation of new knowledge for unmanned aerial vehicle equipment training and the evaluation of new knowledge for unmanned aerial vehicles. The present invention constructs an intelligent unmanned aerial vehicle equipment training system, adjusts the system through a large language model and incremental learning, and realizes intelligent training.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle equipment training, and in particular to a unmanned aerial vehicle equipment training method based on large models and incremental learning. Background Art

[0002] At present, the research on equipment training mainly focuses on equipment training in virtual battlefields, large-scale weapons and equipment training, and equipment training in the field of medical education. However, there are relatively few studies on the rapidly developing and diverse UAV equipment training systems. UAV equipment training plays an important role in promoting the development of the low-altitude economy. In the process of UAV equipment training, there are the following problems:

[0003] (I) The content of test questions is updated slowly and in an untimely manner: The development of UAV equipment is changing with each passing day, and the training system is slow to update the operation training of new UAVs. It relies more on manual updates and lacks automated equipment knowledge base updates, which leads to the inability to provide timely and effective training services in emergency situations.

[0004] (II) The training mode is relatively simple: some UAV equipment training remains in the manual mode stage, such as "master and apprentice" and "blackboard and chalk". There is almost no modern and intelligent training mode, the training method is simple and lacks innovation.

[0005] (III) Lack of personalized training methods for students of different levels: Currently, the training plan, training content, and training result evaluation for UAV equipment are still unified. However, in fact, different students have different learning abilities, past experiences, positions, etc., so different and personalized training questions are needed to improve the training effect. Summary of the invention

[0006] The purpose of the present invention is to provide a method for training unmanned aerial vehicle equipment based on large models and incremental learning, including (1) intelligent generation of a knowledge base for training unmanned aerial vehicle equipment. In the face of complex unmanned aerial vehicle equipment data information, it is necessary to establish a knowledge base for training unmanned aerial vehicle equipment to effectively manage information resources. Use a large language model to effectively extract knowledge from the original data, and store the extracted knowledge in the knowledge base to provide material for the generation of subsequent training question banks. This part is to realize intelligent management of unmanned aerial vehicle equipment training knowledge. The system must have the functions of training information entry, information extraction, digital storage, training data analysis, and unmanned aerial vehicle equipment training knowledge update, and automatically classify and archive personal training information, training registration, training plan and other information.

[0007] (2) Personalized test question bank recommendations. Students test their knowledge of UAV equipment through test question training. During the test question training process, considering the differences in students’ work experience, learning ability, etc., the training questions should be “tailored” for different students, and targeted and personalized training questions should be recommended to students.

[0008] (3) Automatic update of equipment knowledge base. With the iterative update of UAV equipment and the emergence of new equipment, equipment training data will increase dynamically. The intelligent UAV equipment training system timely updates the equipment training knowledge base by introducing incremental learning. It can identify and delete outdated information from the existing knowledge base, and incrementally generate new knowledge items to help trainees quickly learn new equipment training knowledge.

[0009] To achieve the above-mentioned purpose, the present invention provides a UAV equipment training system based on large models and incremental learning, including a UAV equipment knowledge base generation module, a training exercise recommendation module and a training knowledge update module;

[0010] The UAV equipment knowledge base generation module is constructed through a large language model, which sequentially performs multi-source data collection, UAV equipment knowledge extraction, and UAV equipment knowledge fusion to generate a knowledge base.

[0011] The recommendation module of the UAV equipment training system uses the exercise generator to generate exercises for the knowledge base, and the exercises are combined to form an exercise bank. The exercise bank is used to simulate the learning curve of the trainees and select the training exercise set;

[0012] The UAV equipment knowledge update module adds incremental learning to realize new knowledge generation and new knowledge evaluation in the knowledge base.

[0013] A method for training unmanned aerial vehicle equipment based on a large model and incremental learning, comprising the following steps:

[0014] Step 1: Build a UAV equipment knowledge base generation module through a large language model, and use the UAV equipment knowledge base generation module to generate a knowledge base. The generation process includes multi-source data collection, training knowledge extraction, and training knowledge fusion.

[0015] Step 2: Use the recommendation module of the UAV equipment training system to output the recommended sequence of exercises using the knowledge base and evaluate the learning effect of the students, so as to track the learning status of the students;

[0016] Step 3: Use the UAV equipment knowledge update module to generate new exercises through incremental learning, evaluate the new exercises, and adjust the content of the knowledge base accordingly based on the evaluation results.

[0017] Preferably, in step 1, the process of constructing a UAV equipment knowledge base generation module through a large language model is as follows:

[0018] (1) Determination of professional fields and scopes. Through the integration and analysis of massive data in the field of UAV equipment, the scope of the UAV equipment field ontology constructed includes the following data: ① Basic data, including the model and basic attributes of UAV equipment; ② Business data: fault repair, management and support; ③ Others: principle knowledge, R&D units;

[0019] (2) Define the concept categories of UAV equipment and the relationship between them. The concept categories are the basic framework for establishing the ontology. A top-down approach is adopted to establish the concept categories of UAV equipment. This includes: ① defining core concepts: UAV equipment can be classified according to different standards. According to the flight platform configuration, it can be divided into fixed-wing UAVs, rotary-wing UAVs, unmanned airships, paraglider UAVs and flapping-wing UAVs; defining UAV equipment as the first-level core concept, and defining fixed-wing UAVs, rotary-wing UAVs, unmanned airships, etc. as the second-level core concepts; ② further refining the second-level core concepts to obtain more fine-grained UAV equipment concepts;

[0020] (3) Define the relationship between concept categories. The concept categories of UAV equipment involve general semantic relationships, including ① overlapping relationships; ② part-whole relationships, including the composition of UAV equipment; ③ relationship between concept categories and entities; ④ attribute relationships, that is, the relationship between concept categories and attributes;

[0021] The specific output content of the UAV equipment knowledge base generation module is object attributes, description, definition domain, and value range.

[0022] Preferably, in step 1, the specific process of generating the knowledge base is as follows:

[0023] Multi-source data collection: Collect drone equipment data from drone databases, drone encyclopedias, weapon libraries, equipment instructions, training manuals, conferences, papers, and the Internet;

[0024] Training knowledge extraction: Use the large language model to extract knowledge from drone equipment data. The specific extraction process is as follows:

[0025] Preprocessing of drone equipment data: Use tokenization to segment text collected from multi-source data into tokens, each token corresponds to a word, and word segmentation to split the word into smaller subwords; clean up and delete irrelevant characters and punctuation;

[0026] Data input: The preprocessed data is passed to the tokenizer, which converts the text data into the input format of the BERT model. The processed data is input into the BERT model, which maps the words into a high-dimensional vector space and obtains the word vector representation of the vocabulary.

[0027] UAV equipment relationship extraction: The self-attention mechanism is used to focus on the importance between words. The calculation formula is as follows:

[0028] ;

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] in, are all trainable parameter matrices, where To query the training parameter matrix, is the value training matrix, is the keyword training parameter matrix, is the trainable parameter matrix, is the number of words in the sentence, is the vector dimension, is the dimension after linear transformation, Represents the entire sentence information vector, is the probability that the correct answer is a certain relationship, is the overall simulation entity information vector; is the query matrix, is the keyword matrix, is the value matrix; Avgpool represents average pooling, Softmax and Sigmoid represents the activation function, which normalizes the numerical vector into a probability distribution vector;

[0034] Drone equipment entity extraction: After obtaining the relationship vector, concatenate it with the word vector in the sentence, use the conditional random field to identify the entity, and calculate the label The probability formula is as follows:

[0035] ;

[0036] in, is the fusion value matrix, is the feature fusion function, is the word vector matrix, is the parameter matrix, is the relationship vector, is the probability calculation formula for the label, For label The emission matrix, For Transfer to The transfer matrix, For label The emission matrix, For Transfer to The transfer matrix, is the emission matrix, is the trainable parameter matrix, is the transfer matrix, are possible labels in the sequence, is the number of labels, combining relations and entities to obtain triples;

[0037] Training knowledge fusion: In the process of knowledge fusion, there are many differences between knowledge. The training knowledge is fused using the similarity based on cosine distance and Jacard similarity. The calculation formula is as follows:

[0038] Cosine similarity calculation:

[0039] ;

[0040] In the above formula, n represents the n-dimensional attributes of text vectors A and B, and the Jacard similarity calculation is:

[0041] ;

[0042] A and B are numerical vector representations of text vectors, specifically referring to the extraction results of drone equipment data in the previous process. Then, the similarity between data is analyzed based on cosine similarity, so as to perform data classification and cluster analysis. Knowledge fusion is performed based on the calculation results of similarity, including entity linking, record linking and ontology matching. Entity linking matches entities in text with corresponding entities in knowledge base, record linking identifies records describing the same entity in different data sources, and ontology matching discovers the correspondence between concepts in different knowledge bases. Finally, the fused knowledge is stored in the knowledge base in the form of triples.

[0043] Preferably, in step 2, the working process of the drone equipment training system recommendation module is as follows:

[0044] UAV equipment training exercise generation: link the UAV equipment knowledge base with a mature exercise generator to achieve data interaction; extract relevant knowledge points and concepts from the UAV equipment knowledge base, and input different types of exercises into the back end of the exercise generator based on the extracted knowledge, such as multiple-choice questions, fill-in-the-blank questions, short-answer questions, calculation questions, case analysis questions, etc. The exercise generator generates corresponding training question types based on the input instructions and stores the generated exercises in the exercise database;

[0045] Input layer: input student answer records;

[0046] KCPL layer: Use the LSTM model to predict the knowledge points that should be included in the next exercise training, use the learning curve to track the students' learning effect, and judge the students' performance in the simulation operation. The calculation formula is as follows:

[0047] ;

[0048] in, represents the score or number of wrong questions in the first learning test, n represents the cumulative number of exercises, α represents the learning rate, Represents the score or number of wrong questions of the nth learning test, uses recurrent neural network RNN ​​to predict the coverage of knowledge concepts, and uses deep knowledge tracking to predict the students' mastery of knowledge concepts;

[0049] ESFL layer: Based on the analysis results of the previous layer, the exercises in the exercise database are screened to obtain an exercise set, suitable exercises are found for recommendation, and the simulated annealing algorithm is introduced to further screen the exercises in the exercise set to obtain an exercise recommendation sequence;

[0050] Output layer: Outputs the recommended exercise sequence.

[0051] Preferably, in step 3, the working process of the drone equipment knowledge update module is as follows:

[0052] The UAV equipment training knowledge update module introduces an incremental learning algorithm. When the equipment training knowledge base is updated, the execution steps are as follows:

[0053] (1) Identify and delete outdated knowledge items and corresponding test questions from the previously learned knowledge base;

[0054] The deletion criteria are as follows: ① Expert review: Domain experts regularly review the UAV equipment knowledge base to determine which knowledge is outdated or no longer accurate and delete it; ② Knowledge base usage analysis: Set access thresholds for knowledge items or test questions, and delete them when the access threshold cannot be reached within the set time;

[0055] (2) Generate new knowledge items incrementally, select distance-based novelty to evaluate the new knowledge obtained by incremental learning, set the update threshold, and calculate it as follows:

[0056] ;

[0057] in, A collection of new knowledge, for and When the novelty is lower than the update threshold, the system determines that the new knowledge has a high degree of overlap with the existing knowledge and does not need to be changed; when the novelty is higher than the update threshold, the new knowledge is added to the knowledge base;

[0058] Evaluate new exercises. After learning new knowledge and completing the test questions, the system actively pushes evaluation questionnaires to students, inviting them to rate the completed test questions based on their actual learning and training. It collects user feedback through online evaluation and subjective evaluation. Online evaluation is the process of evaluating real user experience indicators and conversion indicators in the process of delivering new knowledge. The main evaluation indicators used are as follows: Average Click-through Rate (ACTR): The click-through rate is the ratio of the number of times a user clicks on a test question in the question bank to the number of times the test question is generated. The average click-through rate is the average of the click-through rates of all test questions. The formula is as follows:

[0059] ;

[0060] Where T is the total number of clicks and N is the total number of generation;

[0061] The average conversion rate ACVR is the ratio of the number of times a user completes a specific test question in the test library to the number of times the test question is generated. The average conversion rate is the average of the conversion rates of all test questions. The formula is as follows:

[0062] ;

[0063] Where C is the number of times the test question is completed, and N is the total number of times the test question is generated;

[0064] Click-through rate (HR) measures the proportion of questions generated for training that contain questions of interest to users and is used to evaluate the accuracy of recommendations and user satisfaction. The formula is as follows:

[0065] ;

[0066] Where N is the total number of users, hits(i) is the actual interaction value of the i-th user;

[0067] Novelty, the proportion of recommended test questions that the user did not know before, is calculated as follows:

[0068] ;

[0069] in The number of questions in the recommended test that the user did not know before, is the total number of recommended test questions.

[0070] Therefore, the present invention adopts the above-mentioned UAV equipment training method based on large model and incremental learning, which has the following advantages:

[0071] In the present invention, firstly, incremental learning is introduced in the process of creating a knowledge base to realize the automatic updating of the UAV equipment training knowledge base; secondly, personalized test questions and test papers are generated, and students choose the difficulty of training according to their own learning situation and set the combination of test papers according to personal preferences, so as to support students' personalized test adaptive training; thirdly, the user's understanding of training knowledge is tested according to the training evaluation results and the score threshold, so as to generate a targeted UAV equipment training plan.

[0072] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a structural diagram of UAV equipment training based on large model and incremental learning in the present invention. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations. The specific model specifications need to be selected and determined according to the actual specifications of the device, and the specific selection calculation method adopts the existing technology in the field, so it will not be described in detail.

[0075] Example

[0076] like Figure 1 As shown, the present invention provides a UAV equipment training method based on a large model and incremental learning, comprising the following steps:

[0077] Step 1: Construct a UAV equipment knowledge base generation module through a large language model. The process of constructing a UAV equipment knowledge base generation module through a large language model is as follows:

[0078] (1) Determination of professional fields and scopes. Through the integration and analysis of massive data in the field of UAV equipment, the scope of the UAV equipment field ontology constructed includes the following data: ① Basic data, including the model and basic attributes of UAV equipment; ② Business data: fault repair, management and support; ③ Others: principle knowledge, R&D units;

[0079] (2) Define the conceptual categories of UAV equipment and the relationships between them. Concept categories are the basic framework for establishing the ontology. A top-down approach is adopted to establish the concept categories of UAV equipment. This includes: ① defining core concepts: UAV equipment can be classified according to different standards. According to the flight platform configuration, it can be divided into fixed-wing UAVs, rotary-wing UAVs, unmanned airships, paraglider UAVs and flapping-wing UAVs; defining UAV equipment as the first-level core concept, and defining fixed-wing UAVs, rotary-wing UAVs, unmanned airships, etc. as the second-level core concepts; ② further refining the second-level core concepts to obtain more fine-grained UAV equipment concepts; taking fixed-wing UAVs as an example, by analyzing the Chinese Weapons and Equipment Encyclopedia, upper monoplane, middle monoplane, lower monoplane, etc. are defined as the third-level concepts.

[0080] (3) Define the relationship between concept categories. The concept category of UAV equipment involves general semantic relationships, including ① overlapping relationships; for example, if UAV system A performs rescue mission X, then the time and space of A overlap with the time and space of rescue mission X; ② the relationship between part and whole, including the composition of UAV equipment; ③ the relationship between concept category and entity; ④ attribute relationship, that is, the relationship between concept category and attribute;

[0081] The specific output content of the UAV equipment knowledge base generation module is object attributes, description, definition domain, and value range.

[0082] The knowledge base is generated using the UAV equipment knowledge base generation module. The generation process includes multi-source data collection, training knowledge extraction and training knowledge fusion.

[0083] The specific process of generating a knowledge base is as follows:

[0084] Multi-source data collection: collect drone equipment information data from drone databases, drone encyclopedias, weapon libraries, equipment instructions, training manuals, conferences, papers, networks, etc.;

[0085] Training knowledge extraction: Use the large language model to extract knowledge from drone equipment data. The specific extraction process is as follows:

[0086] Preprocessing of drone equipment data: Use tokenization to segment the text collected by multi-source drone data into tokens, each token corresponds to a word, and word segmentation to split the word into smaller subwords; clean up and delete irrelevant characters and punctuation;

[0087] Data input: The preprocessed data is passed to the tokenizer, which converts the text data into the input format of the BERT model. The processed data is input into the BERT model, which maps the words into a high-dimensional vector space and obtains the word vector representation of the vocabulary.

[0088] UAV equipment relationship extraction: The self-attention mechanism is used to focus on the importance between words. The calculation formula is as follows:

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] in, are all trainable parameter matrices, where To query the training parameter matrix, is the value training matrix, is the keyword training parameter matrix, is the trainable parameter matrix, is the number of words in the sentence, is the vector dimension, is the dimension after linear transformation, Represents the entire sentence information vector, is the probability that the correct answer is a certain relationship, is the overall simulation entity information vector; is the query matrix, is the keyword matrix, is the value matrix; Avgpool represents average pooling, Softmax and Sigmoid represents the activation function, which normalizes the numerical vector into a probability distribution vector;

[0095] Drone equipment entity extraction: After obtaining the relationship vector, concatenate it with the word vector in the sentence, use the conditional random field to identify the entity, and calculate the probability formula of the label as follows:

[0096] ;

[0097] in, is the fusion value matrix, is the feature fusion function, is the word vector matrix, is the parameter matrix, is the relationship vector, is the probability calculation formula for the label, For label The emission matrix, For Transfer to The transfer matrix, For label The emission matrix, For Transfer to The transfer matrix, is the emission matrix, is the trainable parameter matrix, is the transfer matrix, are possible labels in the sequence, is the number of labels. Combine relations with entities to obtain triples.

[0098] Training knowledge fusion: In the process of knowledge fusion, there are many differences between knowledge. The training knowledge is fused using the similarity based on cosine distance and Jacard similarity. The calculation formula is as follows:

[0099] Cosine similarity calculation:

[0100] ;

[0101] Jacard similarity calculation:

[0102] ;

[0103] A and B are numerical vector representations of text vectors, specifically referring to the extraction results of drone equipment data in the previous process. Then, the similarity between data is analyzed based on cosine similarity, so as to perform data classification and cluster analysis. Knowledge fusion is performed based on the calculation results of similarity, including entity linking, record linking and ontology matching. Entity linking matches entities in text with corresponding entities in knowledge base, record linking identifies records describing the same entity in different data sources, and ontology matching discovers the correspondence between concepts in different knowledge bases. Finally, the fused knowledge is stored in the knowledge base in the form of triples.

[0104] Step 2: Use the recommendation module of the UAV equipment training system to output the recommended sequence of exercises using the knowledge base and evaluate the learning effect of the students, so as to track the learning status of the students;

[0105] The working process of the training exercise recommendation module is as follows:

[0106] Input layer: input student answer records

[0107] KCPL layer (the Knowledge Concept Prediction Layer): Use the LSTM model to predict the knowledge points that should be included in the next exercise training, use the learning curve to track the learning effect of the students, and judge the performance of the students in the simulation operation. The calculation formula is as follows:

[0108] ;

[0109] in, represents the score or number of wrong questions in the first learning test, n represents the cumulative number of exercises, α represents the learning rate, Represents the score or number of wrong questions in the nth learning test; uses recurrent neural network RNN ​​to predict the coverage of knowledge concepts, and uses deep knowledge tracing (DKT) to predict students' mastery of knowledge concepts;

[0110] ESFL layer (the Exercise Set Filtering Layer): Based on the analysis results of the previous layer, the exercises in the question bank are filtered to obtain the Exercise Set (ES), suitable exercises are found for recommendation, and the simulated annealing algorithm is introduced to further filter the exercises in the Exercise Set ES to increase the diversity of exercises, and finally obtain the exercise recommendation sequence;

[0111] Output layer: Outputs the recommended exercise sequence.

[0112] Step 3: Use the equipment training knowledge update module to generate new exercises through incremental learning. The process is as follows:

[0113] The equipment training knowledge update module introduces an incremental learning algorithm. When the equipment training knowledge base is updated, the execution steps are as follows:

[0114] (1) Identify and delete outdated knowledge items and corresponding test questions from the previously learned knowledge base;

[0115] The deletion criteria are as follows: ① Expert review: Domain experts conduct regular reviews of the UAV equipment knowledge base to determine which knowledge is outdated or no longer accurate; ② Knowledge base usage analysis: Set access thresholds for knowledge items or test questions, and delete them if the access threshold cannot be reached within the set time;

[0116] (2) Generate new knowledge items incrementally, select distance-based novelty to evaluate the new knowledge obtained through incremental learning, set the update threshold, and calculate it as follows:

[0117] ;

[0118] in, A collection of new knowledge, for and When the novelty is lower than the update threshold, the system determines that the new knowledge has a high degree of overlap with the existing knowledge and does not need to be changed; when the novelty is higher than the update threshold, the new knowledge is added to the knowledge base;

[0119] Evaluate new exercises. After learning new knowledge and completing the test questions, the system will actively push evaluation questionnaires to students, inviting them to rate the completed test questions based on their actual learning and training. This includes collecting user feedback through online evaluation and subjective evaluation. Online evaluation is the process of evaluating real user experience indicators and conversion indicators in the process of delivering new knowledge. The main evaluation indicators used are as follows:

[0120] The Average Click-through Rate (ACTR): The click-through rate is the ratio of the number of times a user clicks on a question in the question bank to the number of times the question is generated. The average click-through rate is the average of the click-through rates of all questions. The formula is as follows:

[0121] ;

[0122] in is the total number of clicks, is the total number of times generated.

[0123] The Average Conversion Rate (ACVR) is the ratio of the number of times a user completes a specific test question in the test library to the number of times the test question is generated. The average conversion rate is the average of the conversion rates of all test questions. The formula is as follows:

[0124] ;

[0125] Where C is the total number of transformations, is the total number of generations.

[0126] Click-through rate (HR) measures the proportion of questions generated for training that contain questions of interest to users. It can be used to evaluate the accuracy of recommendations and user satisfaction. The formula is as follows:

[0127] ;

[0128] in is the total number of users, For the The actual interaction value of each user;

[0129] Novelty, the proportion of recommended test questions that the user did not know before, is calculated as follows:

[0130] ;

[0131] in The number of questions in the recommended test that the user did not know before, is the total number of recommended test questions.

[0132] The present invention correspondingly provides a UAV equipment training system based on a large model and incremental learning, including a UAV equipment knowledge base generation module, a training exercise recommendation module and a training knowledge update module;

[0133] The UAV equipment knowledge base generation module is constructed through a large language model, which sequentially collects multi-source UAV data, extracts UAV equipment knowledge, and integrates UAV equipment knowledge to generate a knowledge base.

[0134] The recommendation module of the UAV training system generates exercises based on the knowledge base, and the exercises are combined to form an exercise bank. The exercise bank is used to simulate the learning curve of students and select training exercises.

[0135] The UAV equipment training knowledge update module adds incremental learning to the knowledge base to achieve new knowledge generation and new knowledge evaluation.

[0136] The present invention adopts a UAV equipment training method based on large models and incremental learning, uses a large language model to effectively extract knowledge from raw data, and stores the extracted knowledge in a knowledge base to provide material for the generation of subsequent training question banks. This part is to realize the intelligent management of UAV equipment training knowledge. The system must have the functions of training information entry, information extraction, digital storage, training data analysis, and UAV equipment training knowledge update, automatically classify and archive personal training information, training registration, training plan and other information, and generate exercises through an exercise generator. Exercises are then used to test trainees, and trainees are allowed to form a feedback mechanism for the exercises, iterate on exercises and knowledge, and finally realize a continuously updated UAV equipment knowledge base;

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A UAV equipment training method based on large model and incremental learning, characterized in that: The following steps are involved: Step 1: Build a UAV equipment knowledge base generation module through a large language model, and use the UAV equipment knowledge base generation module to generate a knowledge base. The generation process includes multi-source data collection, training knowledge extraction, and training knowledge fusion. In step 1, the specific process of generating the knowledge base is as follows: (1) Determination of professional fields and scopes. Through the integration and analysis of massive data in the field of UAV equipment, the scope of the UAV equipment field ontology constructed includes the following data: ① Basic data, including the model and basic attributes of UAV equipment; ② Business data: fault repair, management and support; ③ Others: principle knowledge, R&D units; (2) Define the concept categories of UAV equipment and the relationship between them. The concept categories are the basic framework for establishing the ontology. A top-down approach is adopted to establish the concept categories of UAV equipment. This includes: ① defining core concepts: UAV equipment is classified according to different standards. According to the flight platform configuration, it can be divided into fixed-wing UAVs, rotary-wing UAVs, unmanned airships, paraglider UAVs and flapping-wing UAVs; UAV equipment is defined as the first-level core concept, and fixed-wing UAVs, rotary-wing UAVs, unmanned airships, etc. are defined as the second-level core concepts; ② further refine the second-level core concepts to obtain more fine-grained UAV equipment concepts; (3) Define the relationship between concept categories. The concept categories of UAV equipment involve general semantic relationships, including ① overlapping relationships; ② part-whole relationships, including the composition of UAV equipment; ③ relationship between concept categories and entities; ④ attribute relationships, that is, the relationship between concept categories and attributes; The specific output content of the UAV equipment knowledge base generation module is object attributes, description, definition domain, and value range; Multi-source data collection: Collect drone equipment data from drone databases, drone encyclopedias, weapon libraries, equipment instructions, training manuals, conferences, papers, and the Internet; Training knowledge extraction: Use the large language model to extract knowledge from drone equipment data. The specific extraction process is as follows: Preprocessing of drone equipment data: Use tokenization to segment text collected from multiple sources into tokens. Each token corresponds to a word. Word segmentation splits the word into smaller subwords. Clean and delete irrelevant characters and punctuation; Data input: The preprocessed data is passed to the tokenizer, which converts the text data into the input format of the BERT model. The processed data is input into the BERT model, which maps the words into a high-dimensional vector space and obtains the word vector representation of the vocabulary. UAV equipment relationship extraction: The self-attention mechanism is used to focus on the importance between words. The calculation formula is as follows: ； ； ； ； ； in, are all trainable parameter matrices, where To query the training parameter matrix, is the value training matrix, is the keyword training parameter matrix, is the trainable parameter matrix, is the number of words in the sentence, is the vector dimension, is the dimension after linear transformation, Represents the entire sentence information vector, is the probability that the correct answer is a certain relationship, is the overall simulation entity information vector; is the query matrix, is the keyword matrix, is the value matrix; Avgpool represents average pooling, Softmax and Sigmoid represents the activation function, which normalizes the numerical vector into a probability distribution vector; Drone equipment entity extraction: After obtaining the relationship vector, concatenate it with the word vector in the sentence, use the conditional random field to identify the entity, and calculate the label The probability formula is as follows: ； in, is the fusion value matrix, is the feature fusion function, is the word vector matrix, is the parameter matrix, is the relationship vector, is the probability calculation formula for the label, For label The emission matrix, For Transfer to The transfer matrix, For label The emission matrix, For Transfer to The transfer matrix, is the emission matrix, is the trainable parameter matrix, is the transfer matrix, are possible labels in the sequence, is the number of labels, combining relations and entities to obtain triples; Training knowledge fusion: In the process of knowledge fusion, there are many differences between knowledge. The training knowledge is fused using the similarity based on cosine distance and Jacard similarity. The calculation formula is as follows: Cosine similarity calculation: ； In the above formula, n represents the n-dimensional attributes of text vectors A and B, and the Jacard similarity calculation is: ； A and B are numerical vector representations of text vectors, specifically referring to the extraction results of drone equipment data in the previous process. Then, the similarity between data is analyzed based on cosine similarity, so as to perform data classification and cluster analysis. Knowledge fusion is performed based on the calculation results of similarity, including entity linking, record linking and ontology matching. Entity linking matches entities in text with corresponding entities in knowledge base, record linking identifies records describing the same entity in different data sources, and ontology matching discovers the correspondence between concepts in different knowledge bases. Finally, the fused knowledge is stored in the knowledge base in the form of triples. Step 2: Use the recommendation module of the UAV equipment training system to output the recommended sequence of exercises using the knowledge base and evaluate the learning effect of the students, so as to track the learning status of the students; Step 3: Use the UAV equipment knowledge update module to generate new exercises through incremental learning, evaluate the new exercises, and adjust the content of the knowledge base accordingly based on the evaluation results.

2. The method for training unmanned aerial vehicle equipment based on large model and incremental learning according to claim 1, characterized in that: In step 2, the working process of the drone equipment training system recommendation module is as follows: UAV equipment training exercise generation: link the UAV equipment knowledge base with a mature exercise generator to achieve data interaction; extract relevant knowledge points and concepts from the UAV equipment knowledge base, and input different types of exercises into the back end of the exercise generator based on the extracted knowledge, including multiple-choice questions, fill-in-the-blank questions, short-answer questions, calculation questions, and case analysis questions. The exercise generator generates corresponding training question types based on the input instructions and stores the generated exercises in the exercise database; Input layer: input student answer records; KCPL layer: Use the LSTM model to predict the knowledge points that should be included in the next exercise training, use the learning curve to track the students' learning effect, and judge the students' performance in the simulation operation. The calculation formula is as follows: ； in, represents the score or number of wrong questions in the first learning test, n represents the cumulative number of exercises, α represents the learning rate, Represents the score or number of wrong questions of the nth learning test, uses recurrent neural network RNN ​​to predict the coverage of knowledge concepts, and uses deep knowledge tracking to predict the students' mastery of knowledge concepts; ESFL layer: Based on the analysis results of the previous layer, the exercises in the exercise database are screened to obtain an exercise set, suitable exercises are found for recommendation, and the simulated annealing algorithm is introduced to further screen the exercises in the exercise set to obtain an exercise recommendation sequence; Output layer: Outputs the recommended exercise sequence.

3. The method for training unmanned aerial vehicle equipment based on large model and incremental learning according to claim 2, characterized in that: In step 3, the working process of the drone equipment knowledge update module is as follows: The UAV equipment training knowledge update module introduces an incremental learning algorithm. When the equipment training knowledge base is updated, the execution steps are as follows: (1) Identify and delete outdated knowledge items and corresponding test questions from the previously learned knowledge base; The deletion criteria are as follows: ① Expert review: Domain experts regularly review the UAV equipment knowledge base to determine which knowledge is outdated or no longer accurate and delete it; ② Knowledge base usage analysis: Set access thresholds for knowledge items or test questions, and delete them when the access threshold cannot be reached within the set time; (2) Generate new knowledge items incrementally, select distance-based novelty to evaluate the new knowledge obtained by incremental learning, set the update threshold, and calculate it as follows: ； in, A collection of new knowledge, for and When the novelty is lower than the update threshold, the system determines that the new knowledge has a high degree of overlap with the existing knowledge and does not need to be changed; when the novelty is higher than the update threshold, the new knowledge is added to the knowledge base; Evaluate new exercises. After learning new knowledge and completing the test questions, the system actively pushes evaluation questionnaires to students, inviting them to rate the completed test questions based on their actual learning and training. It collects user feedback through online evaluation and subjective evaluation. Online evaluation is the process of evaluating real user experience indicators and conversion indicators in the process of delivering new knowledge. The main evaluation indicators used are as follows: Average Click-through Rate (ACTR): The click-through rate is the ratio of the number of times a user clicks on a test question in the question bank to the number of times the test question is generated. The average click-through rate is the average of the click-through rates of all test questions. The formula is as follows: ； Where T is the total number of clicks and N is the total number of generation; The average conversion rate ACVR is the ratio of the number of times a user completes a specific test question in the test library to the number of times the test question is generated. The average conversion rate is the average of the conversion rates of all test questions. The formula is as follows: ； Where C is the number of times the test question is completed, and N is the total number of times the test question is generated; Click-through rate (HR) measures the proportion of questions generated for training that contain questions of interest to users and is used to evaluate the accuracy of recommendations and user satisfaction. The formula is as follows: ； Where N is the total number of users, hits(i) is the actual interaction value of the i-th user; Novelty, the proportion of recommended test questions that the user did not know before, is calculated as follows: ； in The number of questions in the recommended test that the user did not know before, is the total number of recommended test questions.

Images