Low-altitude flight accident intelligent identification method based on multi-source data fusion

Abstract

The invention provides a low-altitude flight accident intelligent identification method based on multi-source data fusion, and the method comprises the steps: carrying out the encryption processing of flight control data, dynamically mapping environment data through an airspace grid mapping model, carrying out the time-space correlation of the two kinds of data, and obtaining multi-source data; constructing a five-dimensional evidence evaluation system through the information entropy of the multi-source data, and carrying out dynamic weight adjustment on the evaluation system according to a real-time data triggering condition to obtain a causal contribution degree; an accident liability contribution degree is obtained through a weight modeling algorithm; then evidence conflict detection is carried out to obtain recombination credibility; and finally, training a dynamic weight distribution model by recombining credibility, and identifying a low-altitude flight accident to obtain a responsibility division result. According to the method, multi-dimensional data are integrated in real time, a complete flight accident evidence chain is constructed, and a dynamic evaluation and multi-party collaborative management mechanism is introduced, so that a scientific basis is provided for responsibility division, and the responsibility affirmation efficiency and credibility are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of low-altitude flight safety management, and particularly relates to a low-altitude flight accident intelligent identification method based on multi-source data fusion. BACKGROUND

[0002] In the field of low-altitude aircraft operation safety, responsibility division is a key link to ensure the effectiveness of airspace management. Due to the involvement of complex airspace environment, diversified participants and dynamic operating conditions in low-altitude flight, responsibility division needs to consider multiple factors such as airspace management rules, aircraft technical performance, operation specifications and external environment to achieve scientific, fair and efficient responsibility identification.

[0003] At present, the responsibility division of low-altitude aircraft mainly relies on manual audit and post-tracing mechanism, which has the defects of low efficiency and strong subjectivity in the process of responsibility identification. For example, in the processing of unmanned aerial vehicle accidents, the responsibility subject may involve operators, manufacturers and third-party service providers, but the existing legal framework lacks clear responsibility division standards, resulting in complex responsibility identification process and frequent disputes. In addition, flight data is usually stored in different systems, lacking a unified data fusion mechanism, making it difficult to build a complete responsibility evidence chain, further reducing the accuracy and reliability of responsibility division.

[0004] The existing supervision technology is mainly based on automatic dependent surveillance-broadcast (ADS-B) and radar monitoring system, but such technology has limited identification ability for unmanned aerial vehicles, and there is a significant supervision blind area. At the same time, although the traditional flight plan reporting and static risk assessment mode can realize the basic supervision function, it cannot respond to dynamic airspace conflicts in real time, and it is difficult to adapt to the increasingly complex operation demand of low-altitude airspace. Therefore, it is necessary to design a low-altitude flight accident intelligent identification method based on multi-source data fusion. SUMMARY

[0005] The purpose of the present application is to provide a low-altitude flight accident intelligent identification method based on multi-source data fusion, to integrate multi-dimensional data in real time, build a complete flight accident evidence chain, and introduce a dynamic evaluation and multi-party collaborative management mechanism, to provide a scientific basis for responsibility division, and improve the efficiency and public credibility of responsibility identification.

[0006] To achieve the above purpose, the present application provides the following scheme:

[0007] A low-altitude flight accident intelligent identification method based on multi-source data fusion, comprising the following steps:

[0008] Collecting flight control data and encrypting and integrating the flight control data through an encryption algorithm to obtain first data;

[0009] The environment data is dynamically mapped through a pre-constructed airspace grid mapping model to obtain second data;

[0010] After verifying the consistency of the first data and the second data, the first data and the second data are spatio-temporally associated to obtain multi-source data;

[0011] A five-dimensional evidence evaluation system is constructed through the information entropy of the multi-source data to obtain a responsibility score;

[0012] The responsibility score is dynamically weighted according to the real-time data trigger condition to obtain a causal contribution degree;

[0013] Based on the causal contribution degree, an accident responsibility contribution degree is obtained through a weight evolution algorithm;

[0014] The accident responsibility contribution degree is subjected to evidence conflict detection to obtain a reorganization credibility;

[0015] A dynamic weight distribution model is constructed, and the dynamic weight distribution model is trained through the reorganization credibility to obtain a final model;

[0016] The final model is used to identify a low-altitude flight accident to obtain a responsibility division result.

[0017] Optionally, flight control data is collected, and the flight control data is encrypted and integrated through an encryption algorithm to obtain first data, including:

[0018] The collected flight control data is converted into JSON format data, and the JSON format data is subjected to data cleaning; the flight control data includes operator instruction records, flight trajectory data and system state logs;

[0019] The digest of the cleaned JSON format data is calculated through an encryption hash function, and a fixed-length hash value is generated;

[0020] The hash value is signed through asymmetric encryption, and the signature is symmetrically encrypted;

[0021] The hash value, the signature and the timestamp are packaged and integrated through a pre-set smart contract to obtain the first data.

[0022] Optionally, the environment data is dynamically mapped through a pre-constructed airspace grid mapping model to obtain second data, including:

[0023] The target airspace is divided into a three-dimensional grid to obtain an airspace grid mapping model;

[0024] The grid in the airspace grid mapping model is dynamically associated with the environment data of the meteorological monitoring points to obtain the second data.

[0025] Optionally, the first data and the second data are spatio-temporally associated after verifying the consistency of the first data and the second data, and multi-source data is obtained, comprising:

[0026] Hash values of the first data and the second data are calculated respectively;

[0027] A spatial grid mapping model and a mapping table of the meteorological monitoring point are introduced into a preset smart contract, and a new contract is obtained;

[0028] The first data and the second data are spatio-temporally associated through the hash values and the new contract, and multi-source data is obtained.

[0029] Optionally, a five-dimensional evidence evaluation system is constructed through the information entropy of the multi-source data, and a responsibility score is obtained, comprising:

[0030] The information entropy of the multi-source data is calculated through an entropy weight method;

[0031] The information entropy is inverted and normalized to obtain an initial weight;

[0032] A five-dimensional evidence evaluation system is constructed based on the initial weight, and a responsibility score is generated.

[0033] Optionally, the responsibility score is dynamically weighted according to a real-time data trigger condition, and a causal contribution degree is obtained, comprising:

[0034] A wind speed sliding window standard deviation is calculated through meteorological mutation detection;

[0035] A causal reasoning model is constructed through a Bayesian network based on the wind speed sliding window standard deviation; the causal reasoning model is composed of an evidence node, an intermediate node, and a target node;

[0036] The causal contribution degree of the target node is calculated through an expectation maximization algorithm.

[0037] Optionally, based on the causal contribution degree, an accident responsibility contribution degree is obtained through a weight evolution algorithm, comprising:

[0038] The causal contribution degree is evaluated for evidence credibility, and a decay index is added to the obtained high conflict evidence to obtain a dynamic weight;

[0039] The dynamic weight is reduced through a multi-head attention mechanism to obtain a short-term weight;

[0040] The short-term weight is normalized through a softmax function to obtain a normalized weight;

[0041] The normalized weight is mapped to an accident responsibility contribution degree through a multi-layer perception machine.

[0042] Optionally, the accident responsibility contribution degree is detected for evidence conflict to obtain a reorganization credibility, comprising:

[0043] calculate the KL divergence of the multi-source data based on the accident responsibility contribution degree;

[0044] compare the KL divergence with a preset conflict threshold to obtain a conflict determination result and calculate a conflict quality of the conflict determination result;

[0045] when the conflict quality is greater than a preset quality threshold, recombine weights through a D-S evidence theory to obtain recombined evidence;

[0046] fuse the recombined evidence through a Dempster combination rule to obtain recombined credibility.

[0047] Optionally, a dynamic weight distribution model is constructed, and the dynamic weight distribution model is trained through the recombined credibility to obtain a final model, including:

[0048] a historical accident data set is collected, and a reference weight is obtained by annotating the historical accident data set by experts;

[0049] continuous evidence in the historical accident data set is converted into fuzzy membership through fuzzy logic;

[0050] the fuzzy membership is probabilistically fused based on the D-S evidence theory to obtain a fused probability, taking the KL divergence as a conflict factor;

[0051] the dynamic weight distribution model is constructed according to the reference weight;

[0052] the dynamic weight distribution model is trained through a minimum prediction weight method based on the fused probability to obtain the final model.

[0053] Optionally, the dynamic weight distribution model is constructed, and the dynamic weight distribution model is trained through the recombined credibility to obtain the final model, and further including: when a confidence degree of the dynamic weight distribution model is lower than a preset confidence threshold, the reference weight is modified through artificial review, the modified result and the reference weight are integrated as a new training set, and the dynamic weight distribution model is parameter-optimized through an incremental learning mechanism.

[0054] According to the specific embodiments of the present application, the following technical effects are disclosed: the low-altitude flight accident intelligent identification method based on multi-source data fusion provided by the present application, the method comprising: collecting flight control data, and encrypting and integrating the flight control data through an encryption algorithm to obtain first data; dynamically mapping the environmental data through a pre-constructed airspace grid mapping model to obtain second data; after verifying the consistency of the first data and the second data, performing spatio-temporal correlation on the first data and the second data to obtain multi-source data; constructing a five-dimensional evidence evaluation system through the information entropy of the multi-source data to obtain a responsibility score; dynamically adjusting the responsibility score according to a real-time data trigger condition to obtain a causal contribution degree; based on the causal contribution degree, obtaining an accident responsibility contribution degree through a weight evolution algorithm; performing evidence conflict detection on the accident responsibility contribution degree to obtain a reorganization credibility; constructing a dynamic weight distribution model and training the dynamic weight distribution model through the reorganization credibility to obtain a final model; identifying a low-altitude flight accident through the final model to obtain a responsibility division result. The method integrates multi-dimensional data in real time, constructs a complete flight accident evidence chain, and introduces a dynamic evaluation and multi-party collaborative management mechanism, thereby providing a scientific basis for responsibility division and improving the efficiency and public credibility of responsibility identification. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0056] Figure 1 The low-altitude flight accident intelligent identification method flowchart of the present application;

[0057] Figure 2 The low-altitude flight accident responsibility division flowchart of the present application. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0059] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0060] As Figure 1 andFigure 2 As shown, the present application provides a low-altitude flight accident intelligent identification method based on multi-source data fusion, comprising the following steps:

[0061] Step 100: Collecting flight control data and encrypting and integrating the flight control data through an encryption algorithm to obtain first data;

[0062] Specifically, the operator instruction record, flight trajectory data, system state log and other flight control data are collected and converted into JSON format. After data cleaning, an encryption hash function is used to calculate the unique digest of the data and generate a fixed-length hash value. Then the hash value is signed using asymmetric encryption to ensure the source is reliable, and the sensitive data is protected using symmetric encryption. Then the signature and hash value are written into the blockchain through the preset smart contract, and the hash value H, signature σ and timestamp T are packaged into a transaction T x .

[0063] It should be noted that the block timestamp and chain structure are used to achieve permanent data solidification, and finally the signature matching, hash consistency and block timestamp range are verified to ensure the integrity of the data and the authenticity of the time sequence.

[0064] Step 200: Dynamically mapping the environment data through a pre-constructed airspace grid mapping model to obtain second data;

[0065] Specifically, the wind speed, air pressure, rainfall and other meteorological data of the meteorological bureau monitoring point are obtained in real time through the API interface; multi-source sensor data such as radar point cloud and ADS-B signal are integrated, and a dynamic airspace map of the target airspace is constructed through SLAM technology. The static background is used for unmanned aerial vehicle self-positioning and map updating, and the dynamic information is separated and integrated into the space-time map to form a dynamic airspace representation containing the position, speed and predicted path of moving objects, and obstacle detection is realized; the meteorological data and obstacle detection results are integrated into environment data. Then the dynamic airspace map is divided into three-dimensional grids, and each grid element is defined as:

[0066] G ijk =[X i ±ΔX,Y j ±ΔY,Z k ±ΔZ];

[0067] Where ΔX, ΔY and ΔZ are grid resolutions, and (X i ,Y j ,Z k ) is the grid center coordinate. Then each grid is dynamically associated with the nearest weather monitoring point, and the nearest weather monitoring point Wijk is assigned to each grid G ijk The matching rule is: when the encrypted GPS coordinates of the unmanned aerial vehicle are uploaded, the real-time coordinates of the unmanned aerial vehicle are dynamically associated with the nearest weather monitoring point Wijku Y u Z u ) mapped to the grid G ijk , the expression of the mapping process is:

[0068]

[0069] Step 300: After verifying the consistency of the first data and the second data, the first data and the second data are spatio-temporally associated to obtain multi-source data;

[0070] Specifically, after the unmanned aerial vehicle system automatically matches the environmental data corresponding to the grid, the grid division rule and the Merkle root of the monitoring point mapping table are pre-written into the smart contract, the double hash values of the flight control data and the environmental data, the timestamp and the grid ID are jointly stored on the chain through the smart contract, and the spatio-temporal consistency is verified, thereby forming multi-source data.

[0071] It should be noted that the verification efficiency is optimized by pre-storing the grid rule and the Merkle tree, while maintaining the tamper-proofing property of the blockchain, and the accurate spatio-temporal association of the flight control data and the environmental data is realized, thereby providing a credible spatio-temporal cross-verification capability for unmanned aerial vehicle supervision.

[0072] Step 400: Construct a five-dimensional evidence evaluation system through the information entropy of the multi-source data to obtain a responsibility score;

[0073] Specifically, the entropy weight method is used to calculate the information entropy of each evidence dimension in the multi-source data, and the initial weight is generated based on the information entropy inversion to ensure that the weight distribution can objectively reflect the information contribution degree of the evidence, and a five-dimensional evidence evaluation system is constructed to generate a responsibility score.

[0074] Further, the initial weight is normalized, and the expression is: The calculation formula of the responsibility score S is: S = ∑(w i ·F(x i ));

[0075] Wherein, w i is the dynamic weight of evidence type i, and F(x i ) is the characteristic function of evidence i. The operation instruction dimension w1 is configured as the highest weight level, the environmental obstacle w2 and the equipment failure w3 are in the secondary weight band, the meteorological data w4 and the air traffic control w5 form the basic monitoring layer, and the weight distribution follows the decision priority principle of human instruction > system state > environmental factor > real-time monitoring.

[0076] Step 500: According to the real-time data trigger condition, the dynamic weight of the responsibility score is adjusted to obtain the causal contribution degree;

[0077] Specifically, the sliding window standard deviation of wind speed is calculated by meteorological mutation detection, and the calculation formula is:

[0078]

[0079] Where xi represents the wind speed, Represents the average wind speed. If σt>θweather (wind speed mutation threshold), it is determined to be a wind speed mutation. Then, a causal inference model is constructed through the Bayesian network, including the evidence node E i , intermediate nodes and target nodes T. Evidence nodes represent various monitoring data, intermediate nodes represent potential causal relationships, and target nodes correspond to accident responsibility determination results. The network parameters of the model are trained using historical accident data, and the expectation maximization algorithm is used to handle missing data. For each evidence node E i , calculate its causal contribution to the target node, the calculation formula is:

[0080]

[0081] Where P(T|E i ) indicates that in evidence E i The probability of target node T occurring when , Indicates that in evidence E i The probability of the target node T occurring when it does not appear, P(T) represents the prior probability of the accident, which is used to standardize the causal contribution.

[0082] Step 600: Based on the causal contribution, the accident responsibility contribution is obtained through a weight evolution algorithm;

[0083] Specifically, multi-evidence conflict detection is performed through the LSTM network. Specifically, the JS divergence is used to quantify the difference in evidence distribution. When JS is greater than the set threshold, it is determined to be high-conflict evidence, and a decay index is applied to the high-conflict evidence to prevent the conflicting evidence from excessively affecting the fusion result, thereby achieving the credibility assessment of the evidence. The expression for the weight of high-conflict evidence is:

[0084]

[0085] Where λ is the attenuation factor. The multi-head attention mechanism is also used to adjust the weight feature, and the processed causal contribution c i Converted into standardized weights for comprehensive integration. The attention mechanism calculates short-term weights through the triplet (Q, K, V) of evidence features, expressed as:

[0086]

[0087] Get the attention weight matrix W that reflects the instantaneous dependency between evidences attnThen, the features V with high correlation in the evidence features are enhanced: V' = W attn • V; then, the weight normalization is realized by a softmax function, and the expression is:

[0088]

[0089] where k is a gain coefficient; finally, the features after attention correction are input into the softmax layer of the multi-layer perception (MLP), so as to map the high-dimensional features to the accident responsibility contribution degree, and the expression is: w i = Softmax(MLP(V i ')).

[0090] Step 700: detecting the evidence conflict of the accident responsibility contribution degree to obtain a reorganization credibility;

[0091] Specifically, based on the accident responsibility contribution degree, the KL divergence of the multi-source data is calculated to quantify the difference in evidence distribution. This embodiment assumes that the probability distribution of two evidence sources for the same event, unmanned aerial vehicle accident, is as follows: evidence source 1: m1(human error) = 0.7, m2(equipment failure) = 0.2, and m3(Θ) = 0.1. Evidence source 2: m1(equipment failure) = 0.8, m2(weather reason) = 0.1, and m3(Θ) = 0.1. Θ represents uncertainty, i.e., the set of all possibilities.

[0092] Then, the quantitative conflict of the KL divergence is calculated. The KL divergence measures the difference between two probability distributions P (evidence source 1) and Q (evidence source 2) and normalizes it. The calculation formula of the KL divergence is:

[0093]

[0094] Then, the KL divergence is compared with the conflict threshold θ set in advance. If DKL(P||Q) ≥ θ, it is determined that there is a conflict. Taking evidence source 1 as an example, the KL divergence is:

[0095]

[0096] At this time, the KL divergence is maximum, i.e., it is determined that there is a conflict. After the determination, the conflict quality K is calculated, and the calculation formula is:

[0097]

[0098] In this embodiment, K is:

[0099] K = m1(human error)·m2(equipment failure) + m1(human error)·m2(weather reason) = 0.63;

[0100] When K is greater than 0.5, it is confirmed that the evidences conflict, and the evidence distribution credibility weight based on JS divergence is expressed as:

[0101] w2=1-w1;

[0102] Meanwhile, the basic probability assignment (BPA) of the evidence source is adjusted to recombine the weight, and the expression is:

[0103] m i ′(A)=w i ·m i (A),

[0104] Finally, the BPA of the recombined evidence source 1 is combined with the evidence source 2 by Dempster to reduce the conflict impact, and the recombined credibility is obtained, and the expression is:

[0105] Step 800: Construct a dynamic weight distribution model, and train the dynamic weight distribution model by the recombined credibility to obtain a final model;

[0106] Specifically, first, the accident scene is classified, and a judgment rule matrix of a typical accident scene is established. The judgment rule matrix of the embodiment is shown in Table 1:

[0107] Table 1 Judgment rule table

[0108]

[0109] Then a large amount of historical accident data is collected, and the baseline weight of each responsible subject in each accident is labeled by a domain expert to form a labeled training set. Then, the fuzzy boundary of continuous evidence is processed by fuzzy logic, the KL divergence is introduced as a conflict factor, and the weight of continuous evidence is corrected by D-S evidence theory. In the embodiment, the continuous evidence such as wind speed change is converted into fuzzy membership by fuzzy logic to quantify its contribution to different responsible subjects. When there is a significant difference in the joint credibility distribution of multi-source evidence, the fuzzy membership is fused by probability using the KL divergence as a conflict factor to obtain a fusion probability. Then, according to the judgment rule matrix, a dynamic weight distribution model with baseline weight as a supervision signal is constructed. Finally, the dynamic weight distribution model is trained end-to-end by minimizing the KL divergence loss between the predicted weight and the real weight.

[0110] Further, during the training process, when the confidence of the dynamic weight allocation model for a certain historical case is lower than the preset confidence threshold 0.7, an artificial review mechanism is triggered, and the benchmark weight revised by the expert is added to the training set as a new sample, the model parameters are continuously optimized through incremental learning mechanism, and finally an final model that can dynamically adapt to new accident mode and the weight allocation is consistent with the expert judgment logic is obtained.

[0111] Step 900: identifying the low-altitude flight accident through the final model to obtain the responsibility division result.

[0112] The beneficial effects of the present application are as follows:

[0113] 1) By encrypting the flight control data, dynamically mapping the airspace grid of the environment data, and fusing multiple source sensors, a tamper-proof full-dimensional evidence chain is constructed, solving the problem of evidence missing caused by traditional data dispersion;

[0114] 2) A five-dimensional evidence evaluation system is innovatively designed, the initial weight is generated based on the entropy weight method, combined with the Bayesian network causal contribution degree calculation and dynamic weight adjustment mechanism, the subjective bias of responsibility judgment is significantly reduced, and the objectivity is enhanced;

[0115] 3) The D-S evidence theory is used to recombine the conflicting evidence, and the multi-head attention mechanism is used to correct the weight features, realizing the second-level response to the dynamic change of airspace, and improving the adaptability to complex scenes.

[0116] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts between various embodiments can be mutually referred to.

[0117] The principle and implementation mode of the present application are described by applying specific examples, and the above embodiment description is only used to help understand the method and core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In conclusion, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for intelligent identification of low-altitude flight accidents based on multi-source data fusion, characterized in that: The steps include: Collecting flight control data, and encrypting and integrating the flight control data using an encryption algorithm to obtain first data; Dynamically mapping the environmental data through a pre-built spatial grid mapping model to obtain second data; After verifying the consistency of the first data and the second data, the first data and the second data are temporally and spatially associated to obtain multi-source data; A five-dimensional evidence evaluation system is constructed through the information entropy of the multi-source data to obtain a responsibility score; Dynamically adjust the weight of the responsibility score according to the real-time data trigger condition to obtain the causal contribution; Based on the causal contribution, the accident responsibility contribution is obtained through a weight evolution algorithm; Conduct evidence conflict detection on the accident responsibility contribution to obtain reconstruction credibility; Constructing a dynamic weight allocation model, and training the dynamic weight allocation model using the reorganization credibility to obtain a final model; The final model is used to identify low-altitude flight accidents and obtain responsibility division results.

2. The method for intelligent identification of low-altitude flight accidents based on multi-source data fusion according to claim 1 is characterized in that: Collecting flight control data, encrypting and integrating the flight control data using an encryption algorithm to obtain first data, including: Convert the collected flight control data into JSON format data and perform data cleaning on the JSON format data; the flight control data includes: operator command records, flight trajectory data, and system status logs; Calculate the digest of the cleaned JSON format data through a cryptographic hash function and generate a hash value of a fixed length; Signing the hash value using asymmetric encryption, and symmetric encrypting the signature; The hash value, the signature, and the timestamp are packaged and integrated through a preset smart contract to obtain the first data.

3. The intelligent identification method for low-altitude flight accidents based on multi-source data fusion according to claim 1 is characterized in that: Dynamically map the environmental data using a pre-built spatial grid mapping model to obtain the second data, including: Dividing the target airspace into three-dimensional grids to obtain the airspace grid mapping model; Dynamically associate the grids in the airspace grid mapping model with the environmental data of the meteorological monitoring point to obtain the second data.

4. The method for intelligent identification of low-altitude flight accidents based on multi-source data fusion according to claim 3 is characterized in that: After verifying the consistency of the first data and the second data, the first data and the second data are temporally and spatially associated to obtain multi-source data, including: Calculating hash values ​​of the first data and the second data respectively; Introducing the airspace grid mapping model and the mapping table of the meteorological monitoring points into a preset smart contract to obtain a new contract; The first data and the second data are temporally and spatially associated with each other through the hash value and the new contract to obtain the multi-source data.

5. The method for intelligent identification of low-altitude flight accidents based on multi-source data fusion according to claim 1 is characterized in that: A five-dimensional evidence evaluation system is constructed based on the information entropy of the multi-source data to obtain a responsibility score, including: Calculating the information entropy of the multi-source data by using an entropy weight method; Inverting and normalizing the information entropy to obtain an initial weight; The five-dimensional evidence evaluation system is constructed based on the initial weights, and the responsibility score is generated.

6. The method for intelligent identification of low-altitude flight accidents based on multi-source data fusion according to claim 1 is characterized in that: The responsibility score is dynamically weighted according to the real-time data triggering conditions to obtain the causal contribution, including: The standard deviation of the wind speed sliding window is calculated by meteorological mutation detection; Based on the wind speed sliding window standard deviation, a causal inference model is constructed through a Bayesian network; the causal inference model consists of evidence nodes, intermediate nodes and target nodes; The causal contribution of the target node is calculated by an expectation maximization algorithm.

7. The method for intelligent identification of low-altitude flight accidents based on multi-source data fusion according to claim 1, characterized in that: Based on the causal contribution, the accident responsibility contribution is obtained through the weight evolution algorithm, including: Evaluate the credibility of the evidence for the causal contribution, and add a decay index to the obtained high-conflict evidence to obtain a dynamic weight; Reduce the dynamic weight through the multi-head attention mechanism to obtain the short-term weight; Normalizing the short-term weights using a softmax function to obtain normalized weights; The normalized weight is mapped to the accident responsibility contribution through a multi-layer perceptron.

8. The method for intelligent identification of low-altitude flight accidents based on multi-source data fusion according to claim 1 is characterized in that: Conduct evidence conflict detection on the accident responsibility contribution to obtain reconstruction credibility, including: Calculating the KL divergence of the multi-source data based on the accident responsibility contribution; Comparing the KL divergence with a preset conflict threshold to obtain a conflict determination result, and calculating the conflict quality of the conflict determination result; When the conflict quality is greater than a preset quality threshold, weight reorganization is performed using DS evidence theory to obtain reorganized evidence; The recombination evidence is fused using the Dempster combination rule to obtain the recombination credibility.

9. The method for intelligent identification of low-altitude flight accidents based on multi-source data fusion according to claim 1, characterized in that: Constructing a dynamic weight distribution model and training the dynamic weight distribution model using the reorganization credibility to obtain a final model, including: Collect historical accident data sets, and have experts label the responsibility of the historical accident data sets to obtain benchmark weights; Converting the continuous evidence in the historical accident data set into fuzzy membership by fuzzy logic; Based on DS evidence theory, the fuzzy membership is probabilistically fused with KL divergence as a conflict factor to obtain the fusion probability; Constructing the dynamic weight allocation model according to the benchmark weight; Based on the fusion probability, the dynamic weight allocation model is trained by minimizing the predicted weight method to obtain the final model.

10. The intelligent identification method for low-altitude flight accidents based on multi-source data fusion according to claim 9 is characterized in that: Constructing a dynamic weight allocation model and training the dynamic weight allocation model through the reorganization credibility to obtain a final model, further comprising: when the confidence of the dynamic weight allocation model is lower than a preset confidence threshold, correcting the benchmark weight through manual review, integrating the correction result and the benchmark weight into a new training set, and optimizing the parameters of the dynamic weight allocation model through an incremental learning mechanism.

Images