Competitive mechanism dual-channel low-altitude wind shear aviation safety risk causal extraction method based on knowledge integration

Abstract

The invention provides a competition mechanism dual-channel low-altitude wind shear aviation safety risk causal extraction method based on knowledge integration. Comprising the following steps: collecting and analyzing low-altitude wind shear unsafe event data of an aviation system, constructing a BERTopic professional field theme identification model for a low-altitude wind shear unsafe event report of the aviation system, and constructing an intuitive causal identification-logical causal identification competition mechanism dual-channel causal relationship extraction model for simulating human causal reasoning thinking. And constructing a cross-level causal decoupling model for low-altitude wind shear aviation risk cause analysis. In order to solve the problems that historical data cannot be fully utilized and the intelligent level is low in low-altitude wind shear risk identification, a field special risk identification model oriented to the specific situation of low-altitude wind shear is designed; on the basis of the risk evolution law of low-altitude wind shear, a brand-new competitive mechanism double-channel causal relationship extraction model for low-altitude wind shear is designed in combination with intuitive causal reasoning and logical causal reasoning, and the method is used for accurately identifying aviation system risks under low-altitude wind shear and deeply mining risk cause links. Theoretical support is provided for improving the overall safety level of an aviation system.

Description

Technical Field

[0001] The present invention relates to risk assessment and management of unsafe events in the aviation system under the specific situation of low-altitude wind shear, specifically, risk identification and risk cause mining of unsafe event reports in the aviation system under low-altitude wind shear, providing a reference for active safety management. Background Art

[0002] The unsafe incident report of the aviation system is an important material and basis for identifying aviation risks, exploring the causes of risks, and anchoring security loopholes. In-depth research on the unsafe incident report is beneficial for the aviation safety management department to find loopholes and fill gaps, eliminate safety hazards, conduct risk control in advance, and continuously improve the effectiveness of proactive safety management.

[0003] Due to the problems of colloquialism, missing data, and unstructured nature of low-altitude wind shear aviation unsafe events, it is difficult to systematically mine and analyze the historical data of low-altitude wind shear aviation unsafe events, resulting in insufficient, non-specific, and in-depth risk identification and cause analysis for most unsafe events. In order to solve the problems of generalized risk identification and superficial cause analysis for low-altitude wind shear aviation system unsafe events, a dual-channel low-altitude wind shear aviation safety risk causal extraction method based on knowledge integration and competition mechanism is designed to enhance the depth of research on aviation unsafe events under low-altitude wind shear and provide a reference for aviation safety management. Summary of the invention

[0004] In order to solve the problem of insufficient risk identification and cause analysis of low-altitude wind shear aviation unsafe events, a dual-channel low-altitude wind shear aviation safety risk causal extraction method based on knowledge integration and competition mechanism was proposed.

[0005] The present invention proposes a dual-channel low-altitude wind shear aviation safety risk causal extraction method based on knowledge integration and competition mechanism, which includes the following steps:

[0006] Step 1: Collect and analyze data related to unsafe events in low-altitude wind shear aviation systems, and build a highly oriented and high-precision professional field dictionary based on data features;

[0007] Step 2: Build a knowledge-driven optimization topic recognition model for low-altitude wind shear aviation unsafe events, efficiently identify meteorological risk factors and regulatory risk factors related to low-altitude wind shear, and assign strong labels to composite risk factors;

[0008] Step 3: Build a knowledge-driven deep learning intuitive causal identifier that simulates the intuitive causal reasoning of the human brain, and achieve efficient causal identification of low-altitude wind shear aviation risk through the low-altitude wind shear aviation risk special external knowledge introduction module and the intelligent context recognition module;

[0009] Step 4: Construct a "association-review-reasoning" logical causal identifier that simulates the logical causal reasoning of the human brain. For the data of low-altitude wind shear aviation unsafe events, multiple optimizations are performed by integrating graph neural networks, memory enhancement networks, and Bayesian deep learning methods to achieve accurate low-altitude wind shear aviation risk causal identification;

[0010] Step 5: Propose a dual-channel causal relationship extraction model with a competitive mechanism, screen and integrate intuitive causal identification results and logical causal identification results with high reliability, break through the limitations of a single causal identification method, and obtain causal identification results with the highest identification accuracy and the greatest mining depth;

[0011] Step 6: Based on the historical data of low-altitude wind shear aviation unsafe events, a data-driven cross-level causal forest decoupling model is constructed to perform cross-level causal decoupling in view of the complex aviation risks of low-altitude wind shear and the high degree of coupling of risk factors.

[0012] To optimize the above technical solutions, the specific measures taken also include:

[0013] Furthermore, step 1 is specifically as follows:

[0014] Step 1.1, collect data related to unsafe events related to low-altitude wind shear in the aviation system, and perform basic data preprocessing such as data cleaning, formatting, and removal of stop words;

[0015] Step 1.2: Use WordPiece and maximum matching algorithm to construct a domain-specific dictionary for low-altitude wind shear aviation risk as external knowledge introduction to obtain sub-word units that meet the characteristics of low-altitude wind shear aviation unsafe event data;

[0016] Step 1.3: Use the GloVe method to obtain the global word vector matrix.

[0017] Furthermore, step 2 is specifically as follows:

[0018] Step 2.1, using mutual information and KL divergence as evaluation indicators, select the UMAP method with the best dimensionality reduction performance on the low-altitude wind shear unsafe event data for data dimensionality reduction, and use the training data for multiple rounds of iterative learning of model parameters;

[0019] Step 2.2, using DBI and Dunn index as evaluation indicators, select the HDBSCAN method with the best clustering performance on the low-altitude wind shear unsafe event data for clustering, and use the training data for multiple rounds of iterative learning of model parameters;

[0020] Step 2.3: Embed the optimized UMAP and HDBSCAN modules into the BERTopic model, introduce a professional dictionary for low-altitude wind shear aviation risks to achieve knowledge-driven, strengthen the identification of meteorological factors and control factors, and give strong labels to complex risk factors.

[0021] Furthermore, step 3 is specifically as follows:

[0022] Step 3.1: Extract the external causal knowledge base for low-altitude wind shear aviation risk through data mining and data enhancement, and introduce the intuitive causal relationship identifier based on BERT;

[0023] Step 3.2: Set a causal relationship binary classification training task with cross entropy loss as the objective function to train the intuitive causal relationship identifier; set a causal strength ranking training task with marginal loss function as the objective function to train the intuitive causal relationship identifier.

[0024] Step 3.3: In view of the characteristics of low-altitude wind shear aviation unsafe events, a high-intensity identification module is set up to strengthen the identification of meteorological factors and control factors; an emphasis module is set up to give strong labels to complex risk factors.

[0025] Furthermore, step 4 is specifically as follows:

[0026] Step 4.1, construct a dedicated causal relationship graph model based on the characteristics of low-altitude wind shear aviation risk data;

[0027] Step 4.2: Conduct causal association for event pairs <e a , e b > sentence, retrieve the causal logic chain {θ 1 ,θ 2 , ...θ J},in Each is a causal rule describing the connection e a , e b The jth causal logic chain;

[0028] Step 4.3: Review the causal knowledge, obtain the neighboring nodes (including context information) and precedent perception representation of each node in the causal chain, and construct the weight matrix and bias matrix corresponding to the causal relationship;

[0029] Step 4.4: Perform causal reasoning and propose a method to use a matrix to calculate the antecedent perception confidence probability of the causal rule and context. The calculation formula is:

[0030]

[0031] In the formula, and refers to the average representation of the eigenvector, W r Refers to the corresponding weight matrix. Delete the causal chain with confidence lower than 0.9;

[0032] Step 4.5: Propose a method to calculate the importance of the causal logic chain. The calculation formula is:

[0033]

[0034] The top 80% causal logic chains ranked in importance were taken as the causal relationship identification results, and a three-stage paradigm was constructed to simulate the human logical reasoning process, namely, a logical causal relationship identification module of association → review → reasoning.

[0035] Furthermore, step 5 is specifically as follows:

[0036] A competition mechanism is constructed, and a calculation method is proposed to compare and analyze the credibility of causal identification of two causal relationship identifiers on each piece of data. The calculation formula is:

[0037]

[0038] In the formula, and It refers to the intuitive causal relationship identification results and the logical causal relationship identification results, and the causal relationship identification results with the highest credibility are selected as the model output.

[0039] Compared with the prior art, the present invention has the following advantages and effects:

[0040] (1) Aiming at the characteristics of low-altitude wind shear aviation unsafe events, which have many risk factors and strong risk orientation, a professional field dictionary for low-altitude wind shear aviation risk was constructed based on the maximum matching algorithm and WordPiece. The optimized BERTopic model was introduced as external knowledge to construct an aviation risk identification model suitable for low-altitude wind shear. The model focused on identifying meteorological risk factors and regulatory risk factors that are closely related to low-altitude wind shear aviation risk, and gave strong labels to compound risk factors, which improved the pertinence and accuracy of low-altitude wind shear aviation risk identification.

[0041] (2) Based on the evolution of low-altitude wind shear aviation risk, a dual-channel causal identification model was constructed that integrates intuitive causal relationship identification and logical causal relationship identification to simulate the human brain's thinking mode. Based on the characteristics of low-altitude wind shear risk, the risk causal chain was deeply explored, and the accuracy and efficiency of the causal analysis of low-altitude wind shear risk was systematically improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Flowchart of the method for aviation risk identification and risk cause mining under low-altitude wind shear DETAILED DESCRIPTION

[0043] The present invention will now be described in further detail with reference to the accompanying drawings.

[0044] Aiming at the problems of generalization of risk identification and superficial cause analysis of aviation system unsafe events under low-altitude wind shear, a dual-channel low-altitude wind shear aviation safety risk causal extraction method based on knowledge integration and competition mechanism is provided. The specific implementation process can be divided into the following steps: Figure 1 shown.

[0045] 1. Collect and analyze relevant data, and build a highly oriented and accurate professional dictionary based on data features

[0046] (1) Collecting data on aviation unsafe events related to low-altitude wind shear

[0047] Collect low-altitude wind shear-related aviation unsafe event data, including the time, location, design model, consequences, event report, equipment status, meteorological conditions, etc. of the unsafe events, and perform basic data preprocessing such as data cleaning, formatting, and stop word removal.

[0048] (2) Constructing a professional dictionary in the aviation field

[0049] WordPiece and maximum matching algorithm are used to construct a professional dictionary in the aviation field for low-altitude wind shear. The unsafe incident reports are segmented according to the constructed professional dictionary to avoid the problems of excessive segmentation of word meanings and incorrect segmentation of professional vocabulary, and to obtain segmentation units that meet the data characteristics of low-altitude wind shear aviation unsafe incident reports.

[0050] (3) Constructing word vector matrix

[0051] Use the GloVe method to obtain the global word vector matrix to prepare data for word embedding of the risk identification model.

[0052] 2. Construct a knowledge-driven optimization topic identification model for low-altitude wind shear aviation unsafe events

[0053] (1) Data Dimensionality Reduction

[0054] Taking mutual information and KL divergence as evaluation indicators, the dimensionality reduction performance of UMAP, t-SNE and other dimensionality reduction methods on low-altitude wind shear related data was compared. The UMAP method with the best dimensionality reduction performance on low-altitude wind shear related data was selected for data dimensionality reduction. Multiple rounds of iterative learning of model parameters were performed to optimize the data dimensionality reduction module.

[0055] (2) Clustering

[0056] Taking DBI and Dunn index as evaluation indicators, the clustering performance of clustering methods such as HDBSCAN, K-Means, and Affinity Propagation on low-altitude wind shear related data was compared. The HDBSCAN method with the best clustering performance on low-altitude wind shear related data was selected for clustering. Multiple rounds of iterative learning of model parameters were performed to optimize the clustering module.

[0057] (3) Module Embedding and BERTopic Topic Modeling

[0058] The optimized UMAP and HDBSCAN modules are embedded in the BERTopic model, and the constructed low-altitude wind shear aviation risk professional dictionary is introduced to realize knowledge-driven, and a knowledge-driven optimized BERTopic aviation risk identification model suitable for low-altitude wind shear is built to efficiently identify meteorological risk factors and control risk factors closely related to low-altitude wind shear, and give strong labels to compound risk factors.

[0059] 3. Build a knowledge-driven deep learning intuitive causal identifier that simulates the intuitive causal reasoning of the human brain

[0060] (1) Introduction of external knowledge

[0061] Through data mining and data enhancement, an external causal knowledge base for low-altitude wind shear aviation risks is refined, and a BERT-based intuitive causal relationship identifier is introduced.

[0062] (2) Model training and optimization

[0063] A causal relationship binary classification training task with cross entropy loss as the objective function is set to train the intuitive causal relationship identifier. A causal strength ranking training task with marginal loss function as the objective function is set to train the intuitive causal relationship identifier, and the intuitive causal identifier is continuously optimized. According to the characteristics of low-altitude wind shear aviation unsafe events, a high-intensity recognition module is set to strengthen the recognition of meteorological factors and control factors; an emphasis module is set to give strong labels to compound risk factors.

[0064] 4. Construct a "association-review-reasoning" logical causal identifier that simulates the logical causal reasoning of the human brain

[0065] (1) Constructing a causal graph model

[0066] A special causal graph model is constructed based on the characteristics of low-altitude wind shear aviation risk data.

[0067] (2) Causal Relationship

[0068] For events with <e a , e b > sentence, retrieve the causal logic chain {θ 1 ,θ2 , ...θ J},in Every is a causal rule describing the connection e a , e b The jth causal logic chain.

[0069] (3) Review of causal knowledge

[0070] For a given causal logic chain In order to examine whether the logic chain is established in the context involved, the context is incorporated into its logic chain as Get event pair <e a , e b >Precedent-aware representations that include contextual information Construct the weight matrix and bias matrix corresponding to the causal relationship.

[0071] (4) Causal Reasoning

[0072] A method is proposed to calculate the confidence probability of the antecedent perception of the causal rule and context. The calculation formula is: In the formula, and refers to the average representation of the eigenvector, W r Refers to the corresponding weight matrix. Delete the causal chain with a confidence level lower than 0.9; propose a method to calculate the importance of the causal logic chain, the calculation formula is: The causal logic chains ranked in the top 80% in importance are taken as the causal reasoning results.

[0073] (5) Building a logical causal relationship identifier

[0074] Integrate causal association, causal knowledge review and causal reasoning modules to construct a three-stage paradigm to simulate the human logical reasoning process, namely, a logical causal relationship identifier of association → review → reasoning.

[0075] 5. Constructing a dual-channel causal relationship extraction model with a competitive mechanism

[0076] (1) Design competition mechanism

[0077] In order to break through the limitations of a single causal identification method, a calculation method for comparing the credibility of the identification results of two causal relationship identifiers is proposed. The calculation formula is: In the formula, and It refers to the intuitive causal relationship identification results and the logical causal relationship identification results, and the causal relationship identification results with the highest credibility are selected as the model output.

[0078] (2) Constructing a dual-channel causal relationship extraction model with a competitive mechanism

[0079] An intuitive causal relationship identifier, a logical causal relationship identifier and a competition mechanism are assembled to form a dual-channel causal relationship extraction model of a competition mechanism. In-depth causal mining is performed on low-altitude wind shear aviation risks to obtain causal identification results with the highest recognition accuracy and the greatest mining depth.

[0080] 6. Constructing a data-driven cross-level causal forest decoupling model

[0081] Based on the historical data of low-altitude wind shear aviation unsafe events, a data-driven cross-level causal forest decoupling model is constructed. Based on the random forest model architecture, a cross-level causal forest decoupling model is developed to achieve cross-level decoupling of low-altitude wind shear aviation risks and high degree of coupling of risk factors. Multiple decoupling is performed at the interface of risk levels to obtain a highly interpretable decoupled causal relationship of low-altitude wind shear aviation risks.

Claims

1. A dual-channel low-altitude wind shear aviation safety risk causal extraction method based on knowledge integration and competition mechanism, characterized by The following steps are involved: 1) Collect and analyze data related to unsafe events in low-altitude wind shear aviation systems, and build a highly oriented and high-precision professional field dictionary based on data features; 2) Construct a knowledge-driven optimization topic identification model for low-altitude wind shear aviation unsafe events, efficiently identify meteorological risk factors and regulatory risk factors related to low-altitude wind shear, and assign strong labels to composite risk factors; 3) Construct a knowledge-driven deep learning intuitive causal identifier that simulates the intuitive causal reasoning of the human brain, and achieve efficient causal identification of low-altitude wind shear aviation risk through the low-altitude wind shear aviation risk special external knowledge introduction module and the intelligent context recognition module; 4) Construct a "association-review-reasoning" logical causal identifier that simulates the logical causal reasoning of the human brain. For the data of low-altitude wind shear aviation unsafe events, multiple optimizations are performed by integrating graph neural networks, memory enhancement networks and Bayesian deep learning methods to achieve accurate low-altitude wind shear aviation risk causal identification; 5) Propose a dual-channel causal relationship extraction model with a competitive mechanism, screen and integrate intuitive causal identification results and logical causal identification results with high reliability, break through the limitations of a single causal identification method, and obtain causal identification results with the highest identification accuracy and the greatest mining depth; 6) Based on the historical data of low-altitude wind shear aviation unsafe events, a data-driven cross-level causal forest decoupling model is constructed to perform cross-level causal decoupling in view of the complex aviation risks of low-altitude wind shear and the high degree of coupling of risk factors.

2. The dual-channel causal extraction method for mining the competitive mechanism of aviation risk causes under low-altitude wind shear according to claim 1 is characterized by: Step 1) Collect and analyze data related to low-altitude wind shear aviation system unsafe events, and conduct in-depth mining of low-altitude wind shear related data by using maximum matching algorithm, WordPiece and other methods to construct a professional dictionary in the field of low-altitude wind shear aviation risk with strong orientation and strong data support; introduce the professional field dictionary as external knowledge and use the GloVe method for data processing to obtain sub-word units and global word vector matrices that meet the characteristics of low-altitude wind shear aviation unsafe event data.

3. The dual-channel causal extraction method for mining the competitive mechanism of aviation risk causes under low-altitude wind shear according to claim 1 is characterized by: Step 2) Build an optimized BERTopic model specifically for risk identification of low-altitude wind shear unsafe events. Based on the characteristics of low-altitude wind shear aviation unsafe events, strengthen the identification of meteorological factors and control factors, and assign strong labels to composite risk factors, providing a structured data basis for risk causal identification.

4. The dual-channel causal extraction method for mining the competitive mechanism of aviation risk causes under low-altitude wind shear according to claim 1 is characterized by: Step 3) Aiming at the uncertainty of low-altitude wind shear aviation risk, a knowledge-driven deep learning intuitive causal identifier is constructed to simulate the intuitive causal reasoning of the human brain. Through the low-altitude wind shear aviation risk special external knowledge introduction module and the intelligent context recognition module, the speed of identifying implicit causal relationships is improved; for events with event pairs, the deep learning intuitive causal identifier is used to simulate the intuitive causal reasoning of the human brain. <e a , e b >, use BERT to encode the sentence and get e a , e b The vector average representation h a ,h b and the average vector representation of the context h c ,Will <e a , e b > and its context representation are input into a double-layer perceptron to obtain Contains all reasoning information for intuitive causal reasoning.

5. The dual-channel causal extraction method for mining the competitive mechanism of aviation risk causes under low-altitude wind shear according to claim 1 is characterized by: Step 4) According to the evolution law of low-altitude wind shear aviation risk, a "association-review-reasoning" logical causal identifier is constructed to simulate the logical causal reasoning of the human brain. Based on the data of low-altitude wind shear aviation unsafe events, multiple optimizations are performed by integrating graph neural networks, memory enhancement networks and Bayesian deep learning methods. In addition, innovative confidence calculation methods and causal importance calculation methods are proposed for the characteristics of low-altitude wind shear aviation unsafe event data. The causal identification results are screened: the causal relationship is converted into a causal relationship graph model; for events with event pairs, the causal relationship is transformed ... <e a , e b > sentence, retrieve the causal logic chain {θ 1 ,θ 2 , ...θ J },in Each is a causal rule that describes the connection a ,e b The jth causal logic chain of ; obtain the neighboring nodes and precedent perception representation of each node in the causal chain, and construct the weight matrix and deviation matrix corresponding to the causal relationship; propose a method to calculate the confidence probability of the antecedent perception of the causal rule and context: A method for calculating the importance of causal logic chains is proposed: Filter causal relationship identification results by importance.

6. The dual-channel causal extraction method for mining the competitive mechanism of aviation risk causes under low-altitude wind shear according to claim 1 is characterized by: Step 5) In order to break through the limitations of a single causal identification method and obtain the causal identification results with the highest identification accuracy and the greatest mining depth, a competition mechanism is introduced and a calculation method is proposed to compare the credibility of causal identification of two causal relationship identifiers on each piece of data. Screen and integrate highly reliable intuitive causal identification results and logical causal identification results to greatly improve the model's adaptability to different data and the accuracy of causal relationship identification.

7. The competitive mechanism dual-channel causal extraction method for mining aviation risk causes under low-altitude wind shear according to claim 1 is characterized by: Step 6) Based on the historical data of low-altitude wind shear aviation unsafe events, a data-driven cross-level causal forest decoupling model is constructed. Based on the random forest model architecture, a cross-level causal forest decoupling model is developed to achieve cross-level decoupling of low-altitude wind shear aviation risks, as the risks are complex and the risk factors are highly coupled. Multiple decoupling is performed at the interface of risk levels to obtain a highly interpretable decoupled causal relationship for low-altitude wind shear aviation risks.

Images