A method and system for constructing a hydrological survey knowledge model based on a large model

Abstract

The present application relates to the technical field of hydrological survey skill training, and particularly relates to a method and system for constructing a hydrological survey knowledge model based on a large model. The method comprises the following steps: obtaining multi-source hydrological data and constructing a hydrological knowledge graph; generating confusion samples corresponding to the hydrological knowledge graph and performing adversarial training on the confusion samples; obtaining bidirectional mapping questions according to the adversarial training results and historical questions in the multi-source hydrological data; inputting the bidirectional mapping questions into a question generation interface of a hydrological survey learning platform; collecting answer data and student physiological data corresponding to the bidirectional mapping questions in real time, and updating student knowledge states according to the answer data and the student physiological data; constructing a hydrological survey knowledge model based on the updated student knowledge states; and generating a student recommended learning scheme using the hydrological survey knowledge model. The present application helps to improve the intelligence and practicality of hydrological survey education.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hydrological survey skill training, and particularly relates to a method and system for constructing a hydrological survey knowledge model based on a large model. BACKGROUND

[0002] Hydrological survey technology plays an important role in water resources allocation, flood control and disaster reduction, and ecological environment monitoring. Traditional hydrological survey mainly relies on field measurement, remote sensing image analysis, and numerical simulation methods based on physical laws to observe and analyze hydrological elements such as precipitation, runoff, evaporation, and groundwater. However, although there are rule-based intelligent test paper generation systems and simple machine learning recommendation models in the prior art, it is difficult to convert hydrological professional knowledge (including standards, typical cases, and instrument operation processes) into structured input understandable by large models, and it is also difficult to achieve long-term memory and dynamic update of student knowledge state. SUMMARY

[0003] Therefore, it is necessary to provide a method and system for constructing a hydrological survey knowledge model based on a large model to solve at least one of the above technical problems.

[0004] To achieve the above-mentioned purpose, a method for constructing a hydrological survey knowledge model based on a large model includes the following steps:

[0005] Step S1: acquiring multi-source hydrological data, defining entities and semantic association relationships in the multi-source hydrological data, and constructing a hydrological knowledge graph;

[0006] Step S2: generating a confusion sample corresponding to the hydrological knowledge graph using a pre-set large model, and performing adversarial training on the confusion sample; obtaining a bidirectional mapping question according to the adversarial training result and historical questions in the multi-source hydrological data;

[0007] Step S3: inputting the bidirectional mapping question into a question generation interface of a hydrological survey learning platform; collecting answer data and student physiological data corresponding to the bidirectional mapping question in real time, and updating student knowledge state according to the answer data and student physiological data;

[0008] Step S4: constructing a hydrological survey knowledge model based on the updated student knowledge state; and generating a student recommended learning plan using the hydrological survey knowledge model.

[0009] The application carries out structured processing and semantic modeling on multi-source hydrological data, constructs a hydrological knowledge graph, generates confusion samples using a large model, and forms bidirectional mapping questions combined with historical questions to improve the model's recognition and reasoning ability of knowledge points through adversarial training. Further, the student's answer data and physiological data such as brain waves are collected in real time as observation values to dynamically update the knowledge graph of the student's mastery of each knowledge point, and an individualized learning plan is generated based on the update result. Compared with the prior art, the application not only converts hydrological professional knowledge such as norms, standards, typical cases and instrument operation processes into structured input understandable by a large model, but also enhances the adaptability and generalization ability of hydrological knowledge by introducing confusion samples and bidirectional mapping questions. At the same time, combined with multi-modal student data, the knowledge state is dynamically updated and long-term memory is realized, effectively solving the problems of difficult individualized recommendation and lagging knowledge update in traditional learning systems, thereby significantly improving the intelligence and practicality of hydrological survey education.

[0010] Optionally, the application also provides a large model-based hydrological survey knowledge model construction system for executing the large model-based hydrological survey knowledge model construction method as described above, which comprises:

[0011] A data acquisition module is configured to acquire multi-source hydrological data, define entities and semantic association relationships in the multi-source hydrological data, and thereby construct a hydrological knowledge graph.

[0012] An adversarial training module is configured to generate confusion samples corresponding to the hydrological knowledge graph using a preset large model, and perform adversarial training on the confusion samples; and obtain bidirectional mapping questions according to the adversarial training results and historical questions in the multi-source hydrological data.

[0013] A knowledge state updating module is configured to input the bidirectional mapping questions to a question generation interface of a hydrological survey learning platform; collect answer data and student physiological data corresponding to the bidirectional mapping questions in real time, and update the student's knowledge state according to the answer data and student physiological data.

[0014] A learning plan recommendation module is configured to construct a hydrological survey knowledge model based on the updated student knowledge state; and generate a recommended learning plan for the student using the hydrological survey knowledge model.

[0015] The large model-based hydrological survey knowledge model construction system of the application can implement any one of the large model-based hydrological survey knowledge model construction methods of the application, and is configured to jointly operate and transmit signals between the modules as a medium to complete the large model-based hydrological survey knowledge model construction method. The modules in the system cooperate with each other, thereby improving the intelligence and practicality of hydrological survey education. BRIEF DESCRIPTION OF DRAWINGS

[0016] Other features, objects, and advantages of the application will become more apparent from the following detailed description when read in connection with the following drawings:

[0017] Fig. 1 A flowchart of the steps of the method for constructing a large model-based hydrological survey knowledge model according to the present application;

[0018] Fig. 2 A schematic diagram of the change in wave frequency data in the electroencephalogram signal in the embodiment of the present application;

[0019] Fig. 3 A module diagram of the system for constructing a large model-based hydrological survey knowledge model in the embodiment of the present application;

[0020] The implementation of the object of the present application, functional features and advantages will be further described with reference to the embodiments in conjunction with the drawings. DETAILED DESCRIPTION

[0021] The technical method of the present application will be described in detail below in conjunction with the drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0022] In addition, the drawings are only schematic illustrations of the present application and are not necessarily drawn to scale. The same reference numerals in the drawings represent the same or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. The functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0023] It should be understood that although the terms "first", "second" and the like can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of the exemplary embodiments, a first element can be referred to as a second element, and similarly a second element can be referred to as a first element. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0024] To achieve the above-mentioned object, please refer to Figs. 1 to 3 The present application provides a method for constructing a large model-based hydrological survey knowledge model, comprising the following steps:​

[0025] Step S1: acquiring multi-source hydrological data, defining entities and semantic association relationships in the multi-source hydrological data, and thus constructing a hydrological knowledge graph;

[0026] In some embodiments, structured hydrological data including rainfall, river level, runoff speed, evaporation, etc. can be acquired through a hydrological monitoring sensor and a historical database calling interface, while unstructured data such as hydrological research reports, geological survey archives, and remote sensing image texts are collected. The above data are first labeled with a unified timestamp, and a distributed data cleaning algorithm is used to remove noise data and missing value entries. Then, the unstructured text is segmented and term annotated based on a hydrological professional dictionary, and hydrological knowledge entities such as rivers, lakes, rainfall, evaporation, and aquifers are defined. The semantic association relationships between entities are established by calculating the co-occurrence frequency of entity pairs greater than 0.6 and the dependency syntax relationship strength of entity pairs greater than 0.7. Finally, the above knowledge is organized into a graph structure in the form of triples (entity-relation-entity), and a hydrological knowledge graph is constructed; the knowledge graph contains no less than 20,000 entity nodes and 50,000 semantic edges, and an indexing mechanism is set to support subsequent efficient retrieval and modeling.

[0027] In another embodiment, multi-source hydrological data and the definition of entities and semantic association relationships in the multi-source hydrological data can also be stored in the form of a database. The database identifies the operation step text in the document through OCR, and LayoutLMv3 parses the device parameter table (such as transmission frequency, beam angle) and calibration flowchart; the annotated entities and relationships can be: annotated entities: ADCP (instrument class), calibration steps (instance class), relationships: ADCP-uses device-current meter, calibration step-follows specification-T / CHES 61-2021. The cutting examples can be: cutting the document into semantic blocks such as device installation, parameter setting, and data correction, and generating a 768-dimensional vector for each block; in Milvus, an index is established, and when a student queries ADCP data anomalies, the associated parameter setting errors and device calibration processes are returned.

[0028] Step S2: generating a confusion sample corresponding to the hydrological knowledge graph using a pre-set large model, and performing adversarial training on the confusion sample; acquiring bidirectional mapping questions according to the adversarial training results and historical questions in the multi-source hydrological data;

[0029] In another embodiment, when generating the confusion sample using a preset large model, a text generation model based on the Transformer structure is selected, which contains 12 layers of encoder and 12 layers of decoder, each layer containing a self-attention module and a feedforward neural network module, and has domain adaptation capability through hydrological corpus pre-training. In the process of constructing the confusion sample, first, independent knowledge points and their adjacent semantic units are selected based on the hydrological knowledge graph to form a context set; dependency syntax analysis is performed on the semantic text of the independent knowledge points to extract core terms, and the core terms are aligned with the context set; when the lexical similarity exceeds 0.75 or the semantic embedding cosine similarity exceeds 0.8, the adjacent semantic unit is determined as a confusion candidate point. Subsequently, three types of confusion samples are generated based on the candidate points: the core term is replaced by the candidate point term to obtain the term replacement type confusion sample; the context relationship is exchanged to obtain the relationship reversal type confusion sample; the principal component features of the candidate points are introduced into the semantic text to obtain the semantic interference type confusion sample. Then the confusion sample is input into the adversarial training framework, which adopts the alternating optimization method of generator and discriminator, the generator rewrites the semantics based on the large model, and the discriminator determines the confusion degree of the sample through the convolutional neural network, and the adversarial training is iterated for 20 rounds to enhance the model robustness. According to the adversarial training result, a confusion semantic feature library is generated, and dependency syntax annotation is performed on the historical test questions to extract the knowledge point distribution and determine the semantic framework of the test questions. Then the confusion semantic features are injected into the semantic framework to construct three types of test questions: discrimination test, logical approximation and reasoning rewriting, and the three types of test questions are merged into a confusion test question set in a ratio of 4:3:3. Finally, a bidirectional mapping relationship between the confusion test question set and the knowledge graph is established to form a bidirectional mapping test question.

[0030] In other embodiments, the large model can use the Qwen-14B large model, which is trained using 50,000 hydrological documents, focusing on learning the semantic representation of terms in the fields of water resources, water level-flow relationship, and sediment particle analysis. The historical test question can be input to describe the main error sources of ADCP flow measurement, and the model can output a test question containing a knowledge graph path (flow test → flow meter method → error analysis) and a capability mapping matrix (device operation ability 0.3, data analysis ability 0.4), which is reviewed by experts and included in the training set.

[0031] Step S3: input the bidirectional mapping test question into the test question interface of the hydrological survey learning platform; real-time collection of test question data and student physiological data corresponding to the bidirectional mapping test question, and updating of student knowledge state according to the test question data and student physiological data;

[0032] In another embodiment, after inputting the bidirectional mapping test question into the test question interface of the hydrological survey learning platform, the student's test question data and physiological data are collected in real time. The test question data includes test accuracy, single question time consumption, and eye movement trajectory, with a sampling frequency of 60 Hz; the physiological data is the electroencephalogram signal collected by a portable electroencephalogram device, which focuses on extracting Wave and Wave power spectrum density, and the time window of 500 ms sliding frequency change value. Then use the extended Kalman filter algorithm to update the knowledge graph of each knowledge point node of the student mastery degree, and the mastery degree is expressed as a probability value of 0~1. For knowledge point nodes that have not appeared for more than 30 days, the mastery degree is attenuated according to a decay factor of 0.95. The system also counts the error mode of the answer, and classifies the test types below 70% accuracy as common error modes, and updates the error distribution in the knowledge graph to depict the weaknesses of the students.

[0033] Step S4: Constructing a hydrological survey knowledge model based on the updated student knowledge state; generating a student recommended learning scheme using the hydrological survey knowledge model.

[0034] In another embodiment, when constructing a learning model based on the updated student knowledge state, a reasoning model based on a graph neural network (GNN) is used, which can perform node representation learning in the knowledge graph to capture the upstream and downstream dependency relationships between knowledge points. The model output is a mastery degree vector corresponding to each knowledge point. When generating a student recommended learning scheme using this model, first define the knowledge point difficulty threshold according to the difficulty distribution of the bidirectional mapping test questions, and set the difficulty threshold to 0.6. Then compare the student's mastery degree vector, if the accuracy rate of three consecutive questions is higher than 85%, the corresponding knowledge point mastery degree is increased by 0.1 level; if the single question time exceeds 25 seconds, the knowledge point mastery degree is decreased by 0.1 level. According to the upgraded and downgraded mastery degree, the subsequent learning path is dynamically adjusted, for example, knowledge points with a mastery degree lower than 0.5 are preferentially pushed, or the proportion of related test questions is increased to 40% of the total number of questions. The final recommended learning scheme not only includes the knowledge point learning order, but also includes the number of questions, difficulty distribution, and learning time arrangement, to achieve personalized adaptive training for students.

[0035] Optionally, generating a student recommended learning scheme in step S4 includes:

[0036] Defining the knowledge point difficulty threshold according to the difficulty of the bidirectional mapping test questions;

[0037] In an embodiment, the system first defines the difficulty threshold of the knowledge point according to the existing difficulty parameter of the bidirectional mapping question. Specifically, the bidirectional mapping question has set an initial difficulty value according to the question discrimination and the confusion semantic complexity when it is constructed, and the difficulty value is distributed in the interval of 0 to 1. Using the quantile division method, the questions with difficulty value below 0.3 are classified as low difficulty, and the corresponding knowledge point difficulty threshold is set to the basic level; the questions with difficulty value in the interval of 0.3 to 0.7 are classified as medium difficulty, and the corresponding knowledge point difficulty threshold is the advanced level; the questions with difficulty value above 0.7 are classified as high difficulty, and the corresponding knowledge point difficulty threshold is the challenge level. In this way, the difficulty threshold of the knowledge point can be ensured to come from the statistical characteristics of the questions, rather than being set arbitrarily, which can form a clear hierarchical structure in the knowledge graph.

[0038] Based on the knowledge point difficulty threshold, and using the hydrological survey knowledge model to generate the updated knowledge point mastery degree corresponding to the knowledge state of the student;

[0039] In another embodiment, the hydrological survey knowledge model based on graph neural network is used to generate the updated knowledge state of the student. The model internally contains three layers of graph convolution layers and two layers of fully connected layers, which can perform representation learning on the nodes in the hydrological knowledge graph, and combine the student's answering performance with the upstream and downstream dependency relationship of the knowledge point to output the mastery degree vector corresponding to each knowledge point, with the value range also being 0~1. The knowledge point difficulty threshold is used as the threshold, when the student's mastery degree is lower than the corresponding difficulty threshold, the knowledge point is marked as a node to be strengthened; when the student's mastery degree is higher than the difficulty threshold, it is marked as a mastered node.

[0040] In other embodiments, using the hydrological survey knowledge model, it can be judged whether the question difficulty is suitable for the student according to the updated knowledge state of the student. For example, the student answers the question correctly but Wave frequency shows high cognitive load, meaning that the knowledge point has not been mastered; on the contrary, low load and correct answer indicate that the student's mastery degree of the knowledge point is strong.

[0041] In another embodiment, the cognitive load and Wave changes can dynamically determine whether the question difficulty is suitable for the student: for example, high cognitive load + long time of gaze can deduce that the question difficulty is too high, and then reduce the question complexity or increase the step analysis in subsequent training; low cognitive load + high accuracy can deduce that the question difficulty is moderate or too low, and then the training difficulty can be increased or reasoning type questions can be increased.

[0042] According to the answer accuracy in the answer data, the knowledge point difficulty corresponding to the trainee is corrected. If the accuracy of three consecutive questions in the answer data is greater than 85%, the knowledge point mastery degree corresponding to the trainee is increased by one level, and the upgraded knowledge point mastery degree is obtained; if the time consumption of any question in the answer data is greater than 25 seconds, the knowledge point mastery degree corresponding to the trainee is reduced by one level, and the degraded knowledge point mastery degree is obtained;

[0043] In another embodiment, the knowledge point difficulty is dynamically corrected according to the answer data of the trainee. If the accuracy of three consecutive questions is more than 85%, it is determined that the trainee has a high mastery stability at the knowledge point, so the knowledge point mastery degree is increased by 0.1, and the maximum is not more than 1.0; if the time consumption of a question in a knowledge point exceeds 25 seconds, it is considered that the cognitive load of the trainee is too high or the understanding path has a problem, so the knowledge point mastery degree is reduced by 0.1, and the minimum is not less than 0.0. The above-mentioned increase and decrease adjustment is executed in the unit of knowledge point granularity, not single question granularity, so as to avoid introducing excessive correction due to accidental answer errors or external interference. The adjustment history is also recorded, and if the knowledge point mastery degree decreases more than three times in a week, it is automatically marked as a key difficulty, and the frequency of the knowledge point is increased in the subsequent learning scheme.

[0044] According to the upgraded knowledge point mastery degree and the degraded knowledge point mastery degree, the learning order and question proportion of each trainee corresponding to the knowledge point in the hydrological survey learning platform are adaptively adjusted, and the trainee recommended learning scheme is obtained.

[0045] In another embodiment, the bidirectional mapping question is adaptively adjusted according to the upgraded mastery degree and the degraded mastery degree. Specifically, in the hydrological survey learning platform, the question bank of each trainee is dynamically generated, and the learning order and question proportion of the knowledge point are determined according to the latest mastery degree vector. When the mastery degree of a knowledge point is increased to more than 0.8, the subsequent question proportion of the knowledge point will automatically decrease to 15% of the total number of questions, so as to avoid redundant training; when the mastery degree of a knowledge point is reduced to less than 0.4, the question proportion of the knowledge point will be increased to 35%, and the related upstream and downstream knowledge points of the knowledge point will also be increased by 5% of the proportion, so as to strengthen the associated learning. The finally generated trainee recommended learning scheme includes: knowledge point sorting, question quantity distribution corresponding to each knowledge point, question difficulty distribution ratio (low difficulty about 30%, medium difficulty about 40%, high difficulty about 30%), and estimated completion time. In this way, the learning platform can provide personalized hydrological survey training path for different trainees, and realize continuous optimization of the knowledge point mastery degree.

[0046] Optionally, step S1 comprises:

[0047] Step S11: Obtain multi-source hydrological data, and define knowledge entities according to the multi-source hydrological data after structured processing;

[0048] In an embodiment, first, raw data is obtained from multi-source hydrological data, which includes real-time flow data of hydrological monitoring stations, geological survey reports, remote sensing images, and meteorological observation data. In order to ensure that the data can be used for subsequent knowledge extraction, the ETL (Extract-Transform-Load) process is used to structure the data, which specifically includes performing sentence segmentation and denoising preprocessing on text data, performing normalization processing on time series numerical data, and performing semantic segmentation on remote sensing images to extract water boundary information. After the structured processing is completed, based on the hydrological professional dictionary, the text corpus is subjected to term recognition, and the monitoring station number, flow index, and geological structure type are defined as knowledge entities. Each entity is stored in the Neo4j graph database through a unique ID code as a basic node of the knowledge graph.

[0049] Step S12: According to the co-occurrence frequency and semantic dependency relationship of each knowledge entity in the multi-source hydrological data, the semantic association relationship of the knowledge entity is defined, and the initial hydrological knowledge graph is constructed based on the semantic association relationship of the knowledge entity and the knowledge entity.

[0050] In another embodiment, the co-occurrence frequency of knowledge entities in the same data source or across data sources is calculated, and the semantic dependency relationship between entities is identified in combination with dependency syntax analysis. For example, there is a semantic dependency relationship between rainfall increase and runoff increase in the hydrological monitoring report, and a edge with a causal attribute is established in the knowledge graph. The point mutual information (PMI) method is used to calculate the co-occurrence frequency threshold. Only when the PMI value of two entities is greater than 0.5, it is considered that they have statistically significant association. At the same time, the dependency syntax tree is introduced, and the verbs “cause” and “affect” are used as semantic labels of the edge, so as to establish the semantic association relationship between entities. Finally, the entity and its semantic association relationship jointly constitute the initial hydrological knowledge graph, in which the node represents the hydrological entity, and the edge represents the semantic dependency of different types.

[0051] Step S13: The multi-source hydrological data is divided into blocks according to the granularity of each knowledge point, and the block result is optimized in combination with the preset hydrological professional dictionary, to obtain a set of hydrological knowledge blocks.

[0052] In another embodiment, the structured multi-source hydrological data is divided into knowledge point granularity. First, sentence and paragraph segmentation is performed on the text corpus, and the number of hydrological terms contained in the sentence and the semantic dependency path coverage are detected using a hydrological professional dictionary. When the coverage rate is above 60%, the sentence is considered as a candidate knowledge point. Subsequently, the similarity value of adjacent sentences is calculated based on the word vector model, and if the similarity is lower than 0.4, the candidate sentence is divided into an independent knowledge point to ensure that the divided knowledge points have distinction and independence. Further, the number of terms of each independent knowledge point is taken as a complexity indicator, and combined with the similarity value to generate a knowledge point granularity score, which ranges from 1 to 5 points, and the higher the score, the finer the granularity of the knowledge point.

[0053] Step S14: Taking the initial hydrological knowledge graph as the index structure and the hydrological knowledge block set as the indexed content carrier, a hydrological knowledge graph is constructed.

[0054] In another embodiment, the initial hydrological knowledge graph is taken as the index structure, and the hydrological knowledge block set is taken as the indexed content carrier, to construct a complete hydrological knowledge graph. Specifically, each knowledge block is bound to the corresponding entity node in the graph, for example, the knowledge block of river flow variation law is related to the entity nodes of river, flow and variation. An inverted index mechanism is used to ensure the bidirectional mapping of knowledge blocks and entities, so that users can retrieve related knowledge blocks through entity nodes and quickly locate related entity nodes through knowledge blocks when querying the knowledge graph. At the same time, a weight value is set for the mapping relationship between the knowledge block and the entity, and the weight is determined according to the semantic similarity between the knowledge block and the entity, and the higher the similarity, the greater the weight. The final obtained hydrological knowledge graph not only contains entities and relationships, but also contains an extended semantic layer with knowledge blocks as content, realizing the unification of indexing and content of knowledge.

[0055] Optionally, the method for calculating the knowledge point granularity in step S13 comprises:

[0056] Text sentence and paragraph segmentation processing is performed on the multi-source hydrological data to obtain a plurality of initial semantic units;

[0057] Text sentence and paragraph segmentation processing is performed on the multi-source hydrological data. The sources of the multi-source data include daily reports of hydrological observation stations, hydrogeological investigation reports and water resources utilization planning texts. A sentence segmenter based on rules and statistics is used, the rules part relies on punctuation marks (such as period, semicolon) as the boundary points, and the statistics part uses a BiLSTM-CRF model to predict potential semantic breakpoints. In the experimental setup, the sentence length is controlled between 15-30 words, and the maximum length of the paragraph is set to 150 words. The segmentation result is a plurality of initial semantic units, which are saved in a structured data table as input for subsequent knowledge point extraction.

[0058] The number of hydrological terms in each initial semantic unit and the semantic dependency relationship between the hydrological terms are detected based on a hydrological professional dictionary. If the dependency path coverage of the semantic dependency relationship between the hydrological terms in the semantic unit reaches a preset coverage threshold, the semantic unit is taken as a candidate knowledge point.

[0059] In another embodiment, the number of hydrological terms in each initial semantic unit and the semantic dependency relationship thereof are detected based on a hydrological professional dictionary. The dictionary contains about 15,000 terms that are manually checked and cover the fields of hydrological processes, geological structures, monitoring indicators, etc. First, the dictionary is used to match and label the terms, and then a dependency syntax analyzer (modified version of Stanford Parser) is used to extract the semantic dependency path. The calculation formula of the dependency path coverage is: coverage=(number of terms with dependency relationship÷total number of terms)×100%. The coverage threshold is set to 60%, that is, when the dependency path coverage between the terms in a semantic unit exceeds 60%, the unit is identified as a candidate knowledge point. For example, in the case of heavy rainfall leading to rapid increase of surface runoff, the terms rainfall and surface runoff establish a causal dependency through the cause, and the coverage exceeds the threshold, so the sentence is marked as a candidate knowledge point.

[0060] The similarity value representation of the candidate knowledge point and its adjacent initial semantic unit is calculated, and if the similarity is lower than a preset similarity threshold, the candidate knowledge point is taken as an independent knowledge point.

[0061] In another embodiment, the similarity value of the candidate knowledge point and its adjacent initial semantic unit is calculated to determine whether the knowledge point has independence. The similarity calculation is based on a double-channel model: on the one hand, Word2Vec is used to generate semantic vectors and calculate the cosine similarity, and on the other hand, the Jaccard coefficient is used to measure the overlap degree of the term set. The two indicators are fused according to the weights of 0.7 and 0.3 to obtain the final similarity score. In practical application, the similarity threshold is set to 0.4, and when the similarity of the candidate knowledge point and the adjacent semantic unit is lower than 0.4, the candidate knowledge point is identified as an independent knowledge point. For example, the fusion similarity between the candidate point “river erosion” and the adjacent semantic unit “water and soil erosion” is only 0.32, so it is identified as an independent knowledge point.

[0062] The number of hydrological terms of the independent knowledge point is taken as the knowledge point complexity index, and the knowledge point granularity is generated by combining the knowledge point complexity index and the similarity value representation.

[0063] In another embodiment, the number of hydrological terms of an independent knowledge point is taken as a knowledge point complexity index, and a knowledge point granularity is generated in combination with the similarity value thereof. The complexity index directly quantifies the number of terms, with a range of 1-10, and the larger the value, the more professional concepts involved in the knowledge point. On this basis, a weighted scoring function is constructed: granularity = a x complexity index + b x (1-similarity value), where a = 0.6 and b = 0.4. This function takes into account both the professional complexity of the knowledge point and the semantic difference from the adjacent unit. Finally, the granularity is divided into five levels, for example, the rainfall intensity is classified as level 1 because it only contains two terms and has a high similarity to the adjacent unit; while the simulation of the runoff process in the watershed is classified as level 5 because it contains eight terms and has a similarity of only 0.25 to the adjacent unit.

[0064] Optionally, generating the confusion sample corresponding to the hydrological knowledge graph in step S2 comprises:

[0065] Based on the hydrological knowledge graph, an independent knowledge point and its adjacent initial semantic unit are selected, and a semantic context set of the independent knowledge point is constructed.

[0066] In one embodiment, based on the hydrological knowledge graph, an independent knowledge point and its adjacent initial semantic unit are selected, and a semantic context set of the independent knowledge point is constructed. Specifically, the knowledge graph consists of three parts: knowledge entities, semantic relationships, and knowledge block indexes. The entity node represents a hydrological professional concept, and the semantic edge represents the dependency or causal relationship. Taking the independent knowledge point as the center, its first-order adjacent nodes are retrieved, and the context window size is limited to three semantic units, thereby obtaining the semantic context set. For example, the adjacent units of the independent knowledge point surface runoff in the graph include rainfall and confluence process, which form a semantic context set together with surface runoff.

[0067] The semantic description text corresponding to the independent knowledge point in the multi-source hydrological data is executed using a large model to perform dependency syntax analysis, and core term extraction is performed according to the syntax analysis result, thereby obtaining the core term of the independent knowledge point.

[0068] In another embodiment, a preset large model is used to perform dependency syntax analysis on the semantic description text corresponding to the independent knowledge points in the multi-source hydrological data, and core terms are extracted according to the analysis results. The large model is a Chinese RoBERTa-wwm-ext structure based on Transformer, and the input is a hydrological description sentence, such as "heavy rainfall leads to an increase in surface runoff". The large model first generates a context vector representation of the input text, and then constructs a syntax dependency tree through a syntax analysis tool (such as LTP or Stanza). The sub-paths centered on causal relationships and modification relationships are filtered on the dependency tree, and the nouns and verbs directly related to the independent knowledge points are extracted as candidate core terms. Secondary verification is performed in combination with a hydrological professional dictionary, and thus the core terms of the independent knowledge points are obtained.

[0069] The semantic context set of the independent knowledge point is aligned with the core term of the independent knowledge point. If the lexical similarity or semantic similarity of the alignment result exceeds the preset corresponding threshold, the adjacent initial semantic unit is taken as a confusion candidate knowledge point of the independent knowledge point.

[0070] In another embodiment, the semantic context set of the independent knowledge point is aligned with the core term thereof, and the lexical similarity and semantic similarity are calculated to screen the confusion candidate knowledge points. The lexical similarity uses the Jaccard coefficient, and the threshold is set to 0.5. The semantic similarity is based on the vector representation generated by the Sentence-BERT model, and the cosine similarity threshold is set to 0.7. When any adjacent semantic unit in the context set and the core term meet the similarity requirement at the lexical or semantic level, the adjacent unit is identified as a confusion candidate knowledge point. For example, the semantic similarity between the core term "runoff" and the adjacent semantic unit "flow measurement" is 0.76, which exceeds the threshold, and therefore "flow measurement" is marked as a confusion candidate knowledge point.

[0071] Based on the semantic description text of the confusion candidate knowledge point in the multi-source hydrological data, multi-type confusion samples are generated.

[0072] In another embodiment, multi-type confusion samples are generated based on the semantic description text of the confusion candidate knowledge point in the multi-source hydrological data. Specifically, sample generation includes three types of strategies: first, a distinction test type, which replaces the core term in the description text with the confusion candidate knowledge point to form a test sample that is similar in semantics but different in answer; second, a logical approximation type, which adjusts the causal relationship or constraint condition in the text, such as changing "rainfall intensity increases leading to an increase in runoff" to "rainfall duration increases leading to an increase in runoff"; and third, a reasoning rewriting type, which replaces the upstream or downstream node in the knowledge point link, such as changing "runoff change" to "sediment content change", thereby introducing reasoning differences. The proportion of the three types of samples can be set to 4:3:3, and the final confusion sample set size is about 120% of the original knowledge point sample.

[0073] Optionally, the multiple types of confusion samples include:

[0074] A term replacement type confusion sample is obtained by replacing a core term in the independent knowledge point with a core term of the confusion candidate knowledge point.

[0075] In an embodiment, a term replacement type confusion sample is obtained by replacing a core term in the independent knowledge point with a core term of the confusion candidate knowledge point. The core term of the independent knowledge point is extracted based on a Transformer combined with a dependency syntax analysis, and the core term of the confusion candidate knowledge point is also obtained by correcting a hydrology professional dictionary. In the generation process, a core term alignment table is first established, and the similarity between the terms is calculated. A word vector is generated using a Word2Vec model, and if the cosine similarity value is greater than 0.6, replacement is allowed. For example, the core term of the independent knowledge point “runoff” is “flow”, and the core term of the confusion candidate knowledge point “flow velocity” is “water flow velocity”, with a similarity of 0.67, which exceeds the threshold. Therefore, “flow” is replaced with “water flow velocity” in the text describing “increased rainfall intensity leading to increased flow”, thereby forming a term replacement type confusion sample.

[0076] A relationship reversal type confusion sample is obtained by exchanging the context relationship between the confusion candidate knowledge point and the independent knowledge point based on the semantic description text of the confusion candidate knowledge point and the independent knowledge point in the multi-source hydrology data.

[0077] In another embodiment, a relationship reversal type confusion sample is obtained by exchanging the context relationship between the confusion candidate knowledge point and the independent knowledge point based on the semantic description text of the confusion candidate knowledge point and the independent knowledge point in the multi-source hydrology data. In implementation, a Sentence-BERT model is used to generate context semantic vectors of the independent knowledge point and the confusion candidate knowledge point, and the directionality of the context dependency edge is detected. If the directionality is opposite on core semantic paths such as causal relationship and modification relationship, exchange is allowed. For example, the dependency path of the independent knowledge point “increased rainfall” and the confusion candidate knowledge point “increased runoff” in the original text is “increased rainfall → increased runoff”, which is exchanged to “increased runoff → increased rainfall”, obtaining a relationship reversal type confusion sample. To prevent generating samples with incomprehensible semantics, the readability score of the exchanged text (calculated using language model perplexity PPL) is required to be no more than 1.2 times that of the original text.

[0078] A semantic interference type confusion sample is obtained by introducing principal component semantic features of the confusion candidate knowledge point into the semantic description text of the independent knowledge point.

[0079] In another embodiment, the semantic interference type confusion sample is obtained by introducing the principal component semantic features of the confusion candidate knowledge point into the semantic description text of the independent knowledge point. The process first reduces the dimension of the description text vector of the confusion candidate knowledge point based on PCA (Principal Component Analysis), extracts the first two principal components as the semantic main features, and inserts them into the semantic description of the independent knowledge point in the form of vector insertion. In operation, linear interpolation is performed with a weight = 0.3, and the formula is , wherein is the semantic vector of the independent knowledge point, is the principal component semantic vector of the confusion candidate knowledge point. Taking "evapotranspiration intensity" as the independent knowledge point and "soil moisture" as the confusion candidate knowledge point as an example, the principal component semantic related to "soil moisture" is injected into the description text "evapotranspiration intensity depends on air temperature and wind speed", so that the text generates "evapotranspiration intensity depends on air temperature, wind speed and potential soil moisture conditions", thereby constructing a semantic interference type confusion sample.

[0080] Especially important is that the proportion of the term replacement type confusion sample, the relationship reversal type confusion sample and the semantic interference type confusion sample in the multi-type confusion sample is 40%, 30% and 30% of the total confusion sample, respectively.

[0081] Optionally, the bidirectional mapping test question obtained in step S2 comprises:

[0082] According to the results of the adversarial training, a confusion semantic feature library is constructed.

[0083] In an embodiment, first, a confusion semantic feature library is constructed based on the results of the adversarial training. Specifically, the adversarial training adopts a combination of the FGSM (Fast Gradient Sign Method) and the PGD (Projected Gradient Descent) of the adversarial sample generation framework to model the semantic disturbance of the confusion samples generated in the hydrological knowledge graph. The RoBERTa-large model is used as a semantic encoder, the original hydrological term and its context description are input, and the disturbance direction in the word vector space is calculated through the adversarial gradient. In the iterative update, the disturbance intensity parameter ε is set to 0.15 to ensure the semantic readability. After completing multiple rounds of adversarial training, the high-frequency disturbance semantic features are extracted, such as the reversal of quantitative indicators and conditional factors, and the replacement of physical quantity boundary conditions, and stored as a confusion semantic feature library. The feature library serves as a disturbance factor for subsequent test question rewriting and provides systematic confusion strategy support for test question generation.

[0084] The historical test question text is extracted from the multi-source hydrological data, the test question text dependency syntax analysis is performed, and the test question knowledge point distribution is labeled according to the test question text dependency syntax analysis result; the test question semantic framework is determined according to the test question knowledge point distribution;

[0085] In another embodiment, historical test texts are extracted from multi-source hydrological data and parsed at the syntax and semantic levels. In a specific implementation, first, dependency syntax analysis is performed on the historical test texts, a dependency syntax analyzer based on a biaffine attention mechanism is used to generate a dependency tree, and a subject-predicate-object structure, a modification relationship, and a causal link are identified in the dependency tree. In combination with a hydrological professional dictionary, the nodes of the dependency tree are labeled, terms related to “hydrological processes”, “hydrological factors”, and “calculation parameters” are marked as candidate knowledge points, and the distribution of test knowledge points is stored in the form of a graph structure. For example, in the test question “calculate the runoff under the rainfall intensity of a certain watershed”, “rainfall intensity” is labeled as an upstream knowledge point, and “runoff” is labeled as a downstream knowledge point, forming an explicit knowledge point link distribution, which is used as the basic structure of the test question semantic framework.

[0086] The confusion semantic feature library is used as a disturbance factor to perform multi-level rewriting on the test question semantic framework to generate bidirectional mapping test questions.

[0087] In another embodiment, the confusion semantic feature library is used as a disturbance factor to perform multi-level rewriting on the test question semantic framework to generate bidirectional mapping test questions. This includes three rewriting strategies: first, term-level rewriting, using replacement terms in the confusion semantic feature library that have a semantic similarity of more than 0.65 with core terms to perform term replacement, thereby generating term substitution type test questions; second, logic-level rewriting, adjusting based on the constraints in the dependency syntax, such as adjusting numerical boundaries or causal order, to ensure that the PPL (Perplexity) of the rewritten text is less than 1.2 times that of the original test question, thereby generating logic approximation type test questions; and third, reasoning chain rewriting, inserting or replacing nodes in the upstream or downstream of the test knowledge point link to construct reasoning expansion type test questions. Finally, the bidirectional mapping test question set is merged according to a preset proportion (term substitution type 40%, logic approximation type 30%, reasoning expansion type 30%) and a bidirectional mapping relationship is established with the original knowledge point link.

[0088] Optionally, the multi-level rewriting of the test question semantic framework includes:

[0089] Replacing or inserting terms in the test question semantic framework with confusion semantic terms in the confusion semantic feature library to construct discrimination test questions;

[0090] In one embodiment, the core terms in the test question semantic framework are replaced or inserted into the confusion semantic feature library to construct a discrimination test question. Specifically, the semantic similarity between the core terms of the test question and the terms in the confusion semantic feature library is calculated using the SimCSE semantic encoding model. If the similarity score is between 0.65 and 0.80, the confusion term is selected as a replacement candidate. To ensure the solvability of the test question, the beam search strategy is used to select the optimal term combination from the candidate set, and perplexity detection (PPL≤120) is performed on the rewritten result to control the language fluency. Through the above replacement operation, the river flow in the original test question can be rewritten as the river section water volume, forming a discrimination test question with semantic interference but logical feasibility.

[0091] Adjust the constraint conditions in the stem of the test question semantic framework to construct a logically approximate test question.

[0092] In another embodiment, based on the stem structure of the test question semantic framework, the constraint conditions in the stem are adjusted to generate a logically approximate test question. Specifically, the condition constraint nodes in the stem syntactic dependency relation tree are identified, such as "rainfall intensity = 30mm / h" and "boundary condition = infiltration rate 0.25". In combination with the disturbance factors in the confusion semantic feature library, the parameters of these constraint conditions are adjusted, with an adjustment range controlled within ±15% to maintain the reasonableness of the test question solution interval. At the same time, the numerical verification module is used to detect whether the solution space of the rewritten test question has a unique solution, and if there are multiple feasible solutions, the adjustment is rolled back. Thus, the original test question "Calculate the runoff of the watershed under the rainfall intensity of 30mm / h" can be rewritten as "Calculate the runoff of the watershed under the rainfall intensity of 35mm / h", forming a logically approximate test question.

[0093] Rewrite the upstream or downstream knowledge points in the knowledge point link in the test question semantic framework to construct a reasoning rewritten test question.

[0094] In another embodiment, the knowledge point link in the test question semantic framework is rewritten to construct a reasoning rewritten test question. Specifically, the upstream and downstream knowledge points corresponding to the target test question are identified using the knowledge point link graph as an index. For example, when the original knowledge point link is "rainfall intensity → runoff", a new adjacent knowledge point "soil moisture content" can be introduced as an upstream knowledge point based on the confusion semantic feature library, or the downstream knowledge point can be replaced with "peak flow". During the rewriting process, the knowledge point dependency consistency verification module is used to ensure that the rewritten link maintains the logical reasonableness, and the link coverage rate calculation (coverage rate ≥ 0.85) is used to verify the integrity of the rewritten knowledge point. This method can effectively improve the test question's examination of the student's reasoning ability and generate a reasoning rewritten test question.

[0095] It is worth noting that the knowledge point dependency consistency checking module is an independent computing unit based on the hydrological knowledge graph, which is used to ensure the causal logic rationality of the knowledge point link after rewriting or expansion. The construction process of the module includes the following steps: first, all knowledge point nodes and their directed dependency edges are extracted from the hydrological knowledge graph to construct a knowledge point dependency matrix, where the matrix row represents the upstream knowledge point, the matrix column represents the downstream knowledge point, and the matrix element takes the dependency strength, ranging from 0 to 1, and the larger the value, the stronger the causal relationship between the knowledge points; second, the module defines a dependency consistency threshold (for example, 0.85) to determine whether the rewritten link maintains the original causal logic; then, when the test question rewriting involves upstream or downstream knowledge point adjustment, the module calculates the coverage rate of the rewritten link in the original knowledge graph through matrix multiplication, coverage rate = total edge weight of rewritten link that meets the original dependency direction / total edge weight of link, if the coverage rate is lower than the threshold, the link will be rolled back or the adjacent knowledge point will be selected again; finally, the module provides a configurable interface to allow adjustment of the dependency strength weight in different disciplines or hydrological scenarios to adapt to knowledge point networks of different complexity.

[0096] The differentiated test questions, the logically approximate test questions and the reasoning rewriting test questions are combined into a confusion test question set according to a preset test question distribution ratio, and a mapping relationship between the confusion test question set and the corresponding knowledge point link is established to obtain a bidirectional mapping test question.

[0097] In another embodiment, the differentiated test questions, the logically approximate test questions and the reasoning rewriting test questions are combined into a confusion test question set according to a preset test question distribution ratio, and a mapping relationship with the knowledge point link is established to obtain a bidirectional mapping test question. Specifically, according to the proportion parameter setting, the differentiated test questions account for 40%, the logically approximate test questions account for 30%, and the reasoning rewriting test questions account for 30%, and each type of test question is stored in the test question library in a hierarchical index. To ensure the effective alignment of the confusion test question set and the knowledge point link, a graph matching-based algorithm is used to compare the knowledge point distribution of each test question with the original link, and if the link matching degree exceeds 0.9, a formal mapping relationship is established, and if it is insufficient, a fine-tuning rewriting is performed. The final bidirectional mapping test question can not only maintain the corresponding relationship with the original knowledge point link, but also have diversified confusion characteristics, which can be used for subsequent student knowledge state evaluation and dynamic learning recommendation.

[0098] Optionally, updating the student knowledge state according to the answer data and the student physiological data in step S3 comprises:

[0099] Real-time collection of answer data corresponding to the bidirectional mapping test question, wherein the answer data includes answer accuracy, time consumption and student eye movement trajectory;

[0100] In an embodiment, the student's answer data for the bidirectional mapping questions is collected in real time by the hydrological survey learning platform. The answer data includes the answer accuracy, the answer time consumption, and the eye movement trajectory. The answer accuracy is recorded in percentage and stored in the answer data matrix immediately after each question is completed; the answer time consumption is in seconds, and the abnormal values are processed by a sliding average window (window length of 3 questions) to form a time series vector; the student's eye movement trajectory is collected by a wearable eye tracking device with a sampling frequency of 120 Hz, and after noise filtering and coordinate standardization, a fixation point matrix and a fixation duration matrix are generated for analyzing the attention distribution and the answer strategy.

[0101] The student's physiological data is collected, wherein the student's physiological data is the wave frequency change data in the electroencephalogram signal;

[0102] In another embodiment, the student's physiological data is collected, and the physiological data is mainly the relative energy change of the wave (8-12 Hz) frequency band in the electroencephalogram signal, which reflects the cognitive load of the student. The electroencephalogram signal is obtained by a head-mounted multi-channel acquisition device with a sampling frequency of 250 Hz. The data is first filtered by a high-pass filter of 0.5 Hz and a low-pass filter of 50 Hz to remove power frequency noise and electromyographic interference, and then the wave power is subjected to short-time Fourier transform to obtain a time series of cognitive load indicators per second, which is further synchronized and matched with the answer data according to the time stamp.

[0103] The student's physiological data and the answer data are taken as observation values, and the student's knowledge state at the previous moment is combined to update the student's mastery degree of each knowledge point node in the hydrological knowledge graph;

[0104] In another embodiment, the collected answer data and the student's electroencephalogram physiological data are taken as observation values to input an extended Kalman filter update model, and the student's knowledge state at the previous moment is combined to update the student's mastery degree of each knowledge point node in the hydrological knowledge graph. The knowledge point node mastery degree is represented by a floating point number between 0 and 1, and each node is updated independently. The model sets the state noise parameter Q=0.01 and the observation noise parameter R=0.05, calculates the node mastery degree through the prediction-update cycle, outputs the updated knowledge state matrix, and reflects the current student's mastery of each knowledge point.

[0105] For the related knowledge point nodes in the hydrological knowledge graph that have not appeared corresponding answer data for more than 30 days, the student's mastery degree of the knowledge point node is reduced according to a preset decay factor; the types of questions with low answer accuracy are counted to update the error mode distribution of the knowledge point node; and the student's knowledge state is updated based on the update state of each knowledge point node in the hydrological knowledge graph.

[0106] Especially important is that the update state of each knowledge point node includes the update state of the mastery degree of the knowledge point node and the update state of the error mode distribution of the knowledge point node.

[0107] In another embodiment, for a knowledge point node in the hydrological knowledge graph for which corresponding answer question data has not appeared for more than 30 days, the system performs exponential decay according to a decay factor of 0.85. The decay formula is , wherein is the number of days after more than 30 days, to ensure that the mastery degree of a long-term review knowledge point decreases reasonably, while providing an initial estimated value for the next answer question, simulating the knowledge forgetting process.

[0108] In another embodiment, the types of test questions with low answer accuracy are counted, and the error mode distribution of the related knowledge point nodes in the hydrological knowledge graph is updated. The error modes can include "concept confusion", "calculation error" and "understanding deviation", etc., recorded in percentage, and the proportion of each type is dynamically adjusted according to the answer records in the last 30 days, forming an error mode set, and stored jointly with the mastery degree matrix of the knowledge point node, to provide a basis for subsequent personalized learning strategies.

[0109] In another embodiment, based on the update state of each knowledge point node in the hydrological knowledge graph, the overall knowledge state of the student is updated synchronously. The update operation integrates the mastery degree of each knowledge point node and the corresponding error mode set, calculates the overall mastery level of the student in the hydrological profession by weighted average, and records the mastery degree change trend, to provide input for subsequent recommended learning scheme generation.

[0110] Optionally, the present application also provides a hydrological survey knowledge model construction system based on a large model, for executing the hydrological survey knowledge model construction method based on a large model as described above, and the hydrological survey knowledge model construction system based on a large model comprises:

[0111] A data acquisition module is configured to acquire multi-source hydrological data, define entities and semantic association relationships in the multi-source hydrological data, and thereby construct a hydrological knowledge graph.

[0112] An adversarial training module is configured to generate confusion samples corresponding to the hydrological knowledge graph by using a preset large model, and perform adversarial training on the confusion samples; and acquire bidirectional mapping test questions according to the adversarial training result and historical test questions in the multi-source hydrological data.

[0113] A knowledge state updating module is configured to input the bidirectional mapping test questions to a test question interface of a hydrological survey learning platform; acquire answer question data and student physiological data corresponding to the bidirectional mapping test questions in real time, and update the knowledge state of the student according to the answer question data and the student physiological data.

[0114] The learning scheme recommendation module is configured to construct a hydrological survey knowledge model based on the updated knowledge state of the trainee, and to generate a recommended learning scheme for the trainee by using the hydrological survey knowledge model.

[0115] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the application being defined by the appended claims and not by the above description, and all changes falling within the meaning and range of equivalency of the elements of the patent file are therefore intended to be embraced within the present application.

[0116] The foregoing is considered as illustrative only of the principles of the application. Numerous modifications and adaptations will be apparent to those skilled in the art in view of the foregoing description. Thus, the embodiments are not to be limited to the examples contained herein but are to be accorded the full scope of the claims, and any and all equivalents thereof, which follow in the appended claims.

Claims

1.A method for constructing a hydrological survey knowledge model based on a large model, characterized by, The method comprises the following steps: Step S1: acquiring multi-source hydrological data, defining entities and semantic association relationships in the multi-source hydrological data, and thus constructing a hydrological knowledge graph; Step S1 comprises: Step S11: acquiring multi-source hydrological data, and defining knowledge entities according to the multi-source hydrological data after structured processing; Step S12: defining semantic association relationships of the knowledge entities according to the co-occurrence frequency and semantic dependency relationships of the knowledge entities in the multi-source hydrological data, and constructing an initial hydrological knowledge graph based on the semantic association relationships of the knowledge entities; Step S13: performing blocking on the multi-source hydrological data according to knowledge point granularity, and combining a preset hydrological professional dictionary to optimize the blocking result, to obtain a hydrological knowledge block set; the calculation method of the knowledge point granularity comprises: performing text sentence and paragraph segmentation processing on the multi-source hydrological data to obtain a plurality of initial semantic units; detecting the number of hydrological terms and the semantic dependency relationships between the hydrological terms in each initial semantic unit based on the hydrological professional dictionary, and if the dependency path coverage of the semantic dependency relationships between the hydrological terms in the semantic unit reaches a preset coverage threshold, regarding the semantic unit as a candidate knowledge point; calculating the similarity value representation of the candidate knowledge point and its adjacent initial semantic units, and if the similarity is lower than a preset similarity threshold, regarding the candidate knowledge point as an independent knowledge point; regarding the number of hydrological terms of the independent knowledge point as a knowledge point complexity index, and combining the knowledge point complexity index and the similarity value representation to generate a knowledge point granularity; Step S14: taking the initial hydrological knowledge graph as an index structure and the hydrological knowledge block set as an indexed content carrier to construct a hydrological knowledge graph; Step S2: generating a confusion sample corresponding to the hydrological knowledge graph by using a preset large model, and performing adversarial training on the confusion sample; acquiring bidirectional mapping test questions according to the adversarial training result and historical test questions in the multi-source hydrological data; Step S3: inputting the bidirectional mapping test questions into a test question interface of a hydrological survey learning platform; collecting answer data and student physiological data corresponding to the bidirectional mapping test questions in real time, and updating student knowledge states according to the answer data and the student physiological data; Step S4: constructing a hydrological survey knowledge model based on the updated student knowledge states; and generating a student recommended learning scheme by using the hydrological survey knowledge model. 2.The method of claim 1, wherein, The generation of the student recommended learning scheme in step S4 comprises: defining a knowledge point difficulty threshold according to the test question difficulty of the bidirectional mapping test questions; generating a knowledge point mastery degree corresponding to the updated student knowledge state based on the knowledge point difficulty threshold and by using the hydrological survey knowledge model; correcting the knowledge point difficulty corresponding to the student according to the answer accuracy in the answer data, if the accuracy of three consecutive questions in the answer data is greater than 85%, the knowledge point mastery degree corresponding to the student is increased by one level to obtain an upgraded knowledge point mastery degree; if the time consumption of any test question in the answer data is greater than 25 seconds, the knowledge point mastery degree corresponding to the student is reduced by one level to obtain a degraded knowledge point mastery degree; According to the upgrade knowledge point mastery degree and the downgrade knowledge point mastery degree, the two-way mapping test questions adaptively adjust the knowledge point learning order and test question proportion of each learner in the hydrological survey learning platform, and obtain a recommended learning scheme of the learner. 3.The method of claim 1, wherein, The confusion sample corresponding to the hydrological knowledge graph generated in step S2 includes: Selecting an independent knowledge point and its adjacent initial semantic unit based on the hydrological knowledge graph, and constructing a semantic context set of the independent knowledge point; Using a large model to perform dependency syntax analysis on the semantic description text of the independent knowledge point in the multi-source hydrological data, and performing core term extraction according to the syntax analysis result to obtain the core term of the independent knowledge point; Aligning the semantic context set of the independent knowledge point with the core term of the independent knowledge point, if the lexical similarity or semantic similarity of the alignment result exceeds the corresponding preset threshold, the adjacent initial semantic unit is taken as a confusion candidate knowledge point of the independent knowledge point; Based on the semantic description text of the confusion candidate knowledge point in the multi-source hydrological data, generate multiple types of confusion samples. 4.The method of claim 3, wherein, The multiple types of confusion samples include: By replacing the core term of the independent knowledge point with the core term of the confusion candidate knowledge point, a term replacement type confusion sample is obtained; Based on the semantic description text of the confusion candidate knowledge point and the independent knowledge point in the multi-source hydrological data, by exchanging the context relationship of the confusion candidate knowledge point and the independent knowledge point, a relationship reversal type confusion sample is obtained; By introducing the principal component semantic features of the confusion candidate knowledge point into the semantic description text of the independent knowledge point, a semantic interference type confusion sample is obtained. 5.The method of claim 1, wherein, The two-way mapping test questions obtained in step S2 include: According to the adversarial training result, construct a confusion semantic feature library; Extract historical test text from multi-source hydrological data, and perform test text dependency syntax analysis, and label the test knowledge point distribution according to the test text dependency syntax analysis result; determine the test semantic framework according to the test knowledge point distribution; Take the confusion semantic feature library as a disturbance factor, and perform multi-level rewriting on the test semantic framework to generate two-way mapping test questions. 6.The method of claim 5, wherein, The multi-level rewriting of the test semantic framework includes: Replace or insert the confusion semantic terms in the confusion semantic feature library into the terms in the test semantic framework to construct a discrimination test question; Adjust the constraint conditions in the stem of the test semantic framework to construct a logical approximation test question; Rewrite the upstream or downstream knowledge points of the knowledge point link in the test semantic framework to construct a reasoning rewriting test question; According to the preset test distribution ratio, the discrimination test question, the logical approximation test question and the reasoning rewriting test question are combined into a confusion test question set, and the mapping relationship between the confusion test question set and the corresponding knowledge point link is established to obtain the two-way mapping test question. 7.The method of claim 1, wherein, In step S3, updating the learner knowledge state according to the answer data and the learner physiological data includes: Real-time collection of answer data corresponding to the two-way mapping test questions, wherein the answer data includes answer accuracy, time consumption and learner eye movement trajectory; Collecting physiological data of the trainee, wherein the physiological data of the trainee is the wave frequency change data in the brain electrical signal ​ Take the learner physiological data and the answer data as observation values, and update the learner mastery degree of each knowledge point node in the hydrological knowledge graph combined with the learner knowledge state obtained at the previous moment; For the related knowledge point node in the hydrological knowledge graph which does not appear corresponding answer question data for more than 30 days, the student mastery degree of the knowledge point node is reduced according to a preset attenuation factor; the types of questions with low answer accuracy are counted to update the error mode distribution of the knowledge point node; and the student knowledge state is updated based on the update state of each knowledge point node in the hydrological knowledge graph. 8.A system for constructing a hydrological survey knowledge model based on a large model, characterized by, The method for constructing a large model-based hydrological survey knowledge model according to claim 1, the large model-based hydrological survey knowledge model construction system comprises: A data acquisition module is configured to acquire multi-source hydrological data, define entities and semantic association relationships in the multi-source hydrological data, and thereby construct a hydrological knowledge graph; An adversarial training module is configured to generate confusion samples corresponding to the hydrological knowledge graph by using a preset large model, and perform adversarial training on the confusion samples; and acquire bidirectional mapping questions according to the adversarial training results and historical questions in the multi-source hydrological data; A knowledge state updating module is configured to input the bidirectional mapping questions into a question generation interface of a hydrological survey learning platform; acquire answer data and student physiological data corresponding to the bidirectional mapping questions in real time, and update the student knowledge state according to the answer data and the student physiological data; A learning scheme recommendation module is configured to construct a hydrological survey knowledge model based on the updated student knowledge state; and generate a recommended learning scheme for the student by using the hydrological survey knowledge model.

Images