Multi-dimensional risk accident tracing method for UAV based on dynamic network model

Abstract

This invention provides a multi-dimensional UAV risk accident tracing method based on a dynamic network model. This method belongs to the technical field of UAV operational risk accident tracing and analysis. The method includes: risk factor identification; construction of a risk evolution network model; model parameter calculation; and risk accident tracing and analysis. This invention aims to clarify the correlations between UAV operational risk factors and conduct tracing and analysis of various UAV risk accidents. This method addresses the lack of timeliness, accuracy, and comprehensiveness in traditional tracing methods.

Description

Technical Field

[0001] The present invention relates to the technical field of tracing and analyzing the risk accidents of unmanned aerial vehicles (UAVs), and in particular to a multi-dimensional risk accident tracing method for UAVs based on a dynamic network model. Background Art

[0002] With the rapid development and widespread application of drone technology, drone operational safety has gradually become a focus of attention. The risk factors associated with drone operations are complex and diverse, involving multiple factors, including operators, maintenance personnel, drone performance, and the operating environment. Existing technologies primarily analyze risk based on statistical results following historical flight accidents. This lacks in-depth analysis of the dynamic correlations between risk factors over time under different environmental influences. This makes it difficult to accurately trace the root causes of accidents in real time, hindering effective prevention and control of drone accidents.

[0003] To address the limitations of existing technologies, the present invention proposes a multi-dimensional UAV risk accident tracing method based on a dynamic network model. Summary of the Invention

[0004] The present invention provides a multi-dimensional UAV risk accident tracing method based on a dynamic network model. This method provides an innovative solution for ensuring the safe and efficient operation of UAVs by comprehensively evaluating multi-dimensional parameters, dynamically analyzing operational risks in real time, and effectively tracing risk propagation paths. It aims to clarify the correlation between UAV operation risk factors and conduct traceability analysis on various types of UAV risk accidents. It is used to solve the problems of lack of timeliness, accuracy and comprehensiveness in traditional tracing methods.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] The multi-dimensional risk accident tracing method of drones based on dynamic network model includes:

[0007] Risk Factor Identification: Analyze historical accident data, flight logs, and real-time operating environment data to identify potential risk factors in drone operations based on the three dimensions of human-machine-environment. Potential risk factors include human factors, drone performance risk factors, and operating environment risk factors.

[0008] Constructing a risk evolution network model: Treating potential risk factors as network nodes and the transmission relationships between risk factors as network edges, constructing a UAV operation risk evolution network model;

[0009] Model parameter calculation: Based on the UAV operation risk evolution network model, the time node degree is introduced to calculate the node degree, the risk propagation intensity is introduced to calculate the closeness centrality, the betweenness centrality is calculated, the dynamic attenuation factor is introduced to calculate the webpage ranking value, and the environmental risk factor is introduced to calculate the edge betweenness;

[0010] Risk accident source tracing analysis: Based on the dynamic changes of risk propagation over time, the risk loss coefficient is introduced, and a variety of network characteristic indicators are comprehensively analyzed to evaluate the importance of nodes, determine the key paths of risk propagation, and extract key risk propagation chains to obtain analysis results.

[0011] In this manual, the formula for calculating node degree by introducing time node degree is as follows:

[0012]

[0013] Where K i (t) represents the time node degree of node i; N represents the total number of network nodes; a ij (t) indicates whether there is a risk propagation edge between nodes i and j at time t. If yes, it takes the value 1; if no, it takes the value 0.

[0014] In this manual, the risk transmission intensity is introduced to calculate the proximity centrality C i The formula is as follows:

[0015]

[0016] Where, d ij The number of edges in the shortest path starting from node i and ending at node j; n represents the total number of nodes; r ij Represents the risk transmission intensity from node i to node j.

[0017] In this specification, the betweenness centrality B is calculated i The formula is as follows:

[0018]

[0019] Where σ(s, t) represents the total number of shortest paths between nodes s and t; σ(s, t|v) represents the number of shortest paths from node s to node t passing through node v.

[0020] In this specification, the formula for calculating the webpage ranking value by introducing the dynamic decay factor d is as follows:

[0021] d=d0+Δd×T(t);

[0022] T(t)=e -λt ;

[0023]

[0024] Where: PR(p i) is the webpage ranking value of the node, n is the total number of nodes; d0 is the initial attenuation factor, Δd is the adjustment amplitude, usually 0.25, T(t) is the time function, λ is the preset constant, usually 0.1; M(p i ) is the same as p j connected nodes; w(p i ) is (p i ,p j )’s edge weight; L(p j ) is p j The degree of the node.

[0025] In this manual, the formula for calculating edge betweenness by introducing environmental risk factors is as follows:

[0026]

[0027] Where E i is the edge betweenness of i; E env (j,k,i) are the environmental risk factors of j, k and i; n jk (i) is the number of paths between node j and node k that pass through edge i.

[0028] In this manual, the risk loss coefficient is introduced and the formula for comprehensively analyzing various network characteristic indicators is as follows:

[0029]

[0030] Where S i is the risk loss function, K i 、C i 、B i PR(p i ), E i are the time node degree, closeness centrality, betweenness centrality, page ranking value, and edge betweenness of node i, respectively; α, β, γ, δ, and μ are the weight coefficients of time node degree, closeness centrality, betweenness centrality, page ranking value, and edge betweenness, respectively, all in the range of [0,1], and α+β+γ+δ+μ=1; is the risk loss coefficient.

[0031] In this manual, the LSTM network is used to predict the time function T(t);

[0032] T(t)=σ(W·h t +b);

[0033] Its hidden state h t The update formula is:

[0034] h t =LSTM(x t-n:t );

[0035] Where σ is the Sigmoid function, W is the weight matrix from the LSTM hidden state to the output, b is the bias term, and x t-n:t is the input feature within the time window.

[0036] In this specification, building a risk evolution network model includes:

[0037] Establish a hierarchical network structure, including the human factors layer, the airborne layer, and the environmental layer;

[0038] Use hypergraph to represent cross-layer risk propagation path, hyperedge weight W cross Calculated by the following formula:

[0039] W cross =εE env +ξT(t);

[0040] ε, ξ are dynamic learning parameters, E env Environmental risk factors.

[0041] In this manual, Kalman filtering is used to fuse multi-sensor data to calculate the real-time environmental risk factor E. env (t):

[0042]

[0043] Among them, s i (t) is the normalized reading of the i-th sensor, T is the temperature parameter, λ i is the normalization coefficient.

[0044] In summary, the present invention has at least the following beneficial effects:

[0045] The present invention can trace the root causes of drone accidents more accurately and in real time, and can be used to prevent and control the occurrence of drone risk accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 This is a schematic diagram of the multi-dimensional risk accident tracing method for drones based on a dynamic network model involved in the present invention.

[0048] Figure 2 Schematic diagram of the UAV operation risk evolution network model involved in the present invention.

[0049] Figure 3 This is a schematic diagram of the hierarchical structure of the UAV operation risk factors involved in the present invention. DETAILED DESCRIPTION

[0050] Hereinafter, only certain exemplary embodiments are briefly described. As will be appreciated by those skilled in the art, the described embodiments may be modified in various ways without departing from the spirit or scope of the embodiments of the present invention. Therefore, the drawings and description are to be regarded as illustrative in nature and not restrictive.

[0051] The disclosure below provides many different embodiments or examples for implementing different structures of the embodiments of the present invention. In order to simplify the disclosure of the embodiments of the present invention, the components and configurations of specific examples are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. In addition, the embodiments of the present invention may repeat reference numerals and / or reference letters in different examples. Such repetition is for the purpose of simplicity and clarity and does not in itself indicate the relationship between the various embodiments and / or configurations discussed.

[0052] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0053] like Figure 1 As shown, this embodiment provides a multi-dimensional UAV risk accident tracing method based on a dynamic network model, including:

[0054] Risk Factor Identification: Analyze historical accident data, flight logs, and real-time operating environment data to identify potential risk factors in drone operations based on the three dimensions of human-machine-environment. Potential risk factors include human factors, drone performance risk factors, and operating environment risk factors.

[0055] Constructing a risk evolution network model: Treating potential risk factors as network nodes and the transmission relationships between risk factors as network edges, constructing a UAV operation risk evolution network model;

[0056] Model parameter calculation: Based on the UAV operation risk evolution network model, the time node degree theory is introduced to calculate the node degree, the risk propagation intensity is introduced to calculate the closeness centrality, the betweenness centrality is calculated, the dynamic attenuation factor is introduced to calculate the webpage ranking value, and the environmental risk factor is introduced to calculate the edge betweenness;

[0057] Risk accident source tracing analysis: Based on the dynamic changes of risk propagation over time, the risk loss coefficient is introduced, and a variety of network characteristic indicators are comprehensively analyzed to evaluate the importance of nodes, determine the key paths of risk propagation, and extract key risk propagation chains to obtain analysis results.

[0058] In some embodiments, the formula for calculating node degree by introducing time node degree is as follows:

[0059]

[0060] Where K i (t) represents the time node degree of node i; N represents the total number of network nodes; a ij (t) indicates whether there is a risk propagation edge between nodes i and j at time t. If yes, it takes the value 1, and if no, it takes the value 0. By using 0 and 1, the node degree calculation process is diverted to reduce the overall calculation workload (generally, calculations are performed on all nodes, and diversion can reduce the calculation of some unrelated nodes).

[0061] In some embodiments, the risk transmission intensity is introduced to calculate the proximity centrality C i The formula is as follows:

[0062]

[0063] Where, d ij The number of edges in the shortest path starting from node i and ending at node j; n represents the total number of nodes; r ij represents the risk transmission intensity from node i to node j; f ij is the number of times the path from node i to j appears in historical accident data. The higher the frequency, the greater the risk transmission intensity.

[0064] Risk transmission intensity r ij It represents the reliability of the risk propagation path from node i to node j in historical data (historical drone accident reports, sensor monitoring logs, simulation test data, etc.), reflecting the stability, repeatability and risk transmission efficiency of the path in risk propagation. ij ×r ij Indicates: For example, the number of edges in the shortest path d ij The less, the risk transmission intensity r ij The larger the value, the faster the risk propagation from node i to j, and the closer it is to the centrality C i The larger the value, the higher the node importance.

[0065] In some embodiments, betweenness centrality B is calculated i The formula is as follows:

[0066]

[0067] Where σ(s, t) represents the total number of shortest paths between nodes s and t; σ(s, t|v) represents the number of shortest paths from node s to node t passing through node v.

[0068] In some embodiments, the formula for calculating the webpage ranking value by introducing the dynamic decay factor d is as follows:

[0069] d=d0+Δd×T(t);

[0070] T(t)=e -λt ;

[0071]

[0072] Where: PR(p i ) is the webpage ranking value of the node, n is the total number of nodes; d0 is the initial attenuation factor, Δd is the adjustment amplitude, T(t) is the time function, λ is a preset constant, usually set to 0.1; M(p i ) is the same as p j connected nodes; w(p i ) is (p i ,p j )’s edge weight; L(p j ) is p j The degree of the node; w(p i ) depends on the risk propagation probability P(i→j) of the edge, that is, the node p i After the occurrence, node p j The conditional probability of this happening is:

[0073]

[0074] Among them, N(i∩j) is the node p i and p j The number of simultaneous occurrences, N(i) is the number of nodes p i The number of separate occurrences.

[0075] In some embodiments, the formula for calculating edge betweenness by introducing environmental risk factors is as follows:

[0076]

[0077] Where E i is the edge betweenness of i; E env (j,k,i) are the environmental risk factors of j, k and i; n jk (i) is the number of paths between node j and node k that pass through edge i.

[0078] In some embodiments, a risk loss coefficient is introduced and the formula for comprehensively analyzing multiple network characteristic indicators is as follows:

[0079]

[0080] Where S i is the risk loss function, K i 、C i 、B i PR(pi ), E i are the time node degree, closeness centrality, betweenness centrality, page ranking value, and edge betweenness of node i, respectively; α, β, γ, δ, and μ are the weight coefficients of time node degree, closeness centrality, betweenness centrality, page ranking value, and edge betweenness, respectively, all in the range of [0,1], and α+β+γ+δ+μ=1; is the risk loss coefficient.

[0081] in, Indicates the comprehensive loss degree that may be caused once a node (risk factor or risk event) occurs or is triggered, including personal injury, property loss, social impact and economic cost. According to the relevant regulations and standards for drones, the accident level is divided into intact, minor damage, major loss and complete loss. Therefore, this invention directly maps the accident level to the risk loss coefficient. In the future, when the amount of historical accident data is sufficient, the accident levels can be further classified and graded to map different risk loss coefficient values.

[0082]

[0083] In some embodiments, an LSTM network is used to predict the time function T(t);

[0084] T(t)=σ(W·h t +b);

[0085] Its hidden state h t The update formula is:

[0086] h t =LSTM(x t-n:t );

[0087] Where σ is the Sigmoid function, W is the weight matrix from the LSTM hidden state to the output, b is the bias term, and x t-n:t is the input feature within the time window.

[0088] In some embodiments, constructing a risk evolution network model includes:

[0089] Establish a hierarchical network structure, including the human factors layer, the airborne layer, and the environmental layer;

[0090] Use hypergraph to represent cross-layer risk propagation path, hyperedge weight W cross Calculated by the following formula:

[0091] W cross =εE env +ξT(t);

[0092] ε, ξ are dynamic learning parameters, E envEnvironmental risk factors.

[0093] In some embodiments, a Kalman filter is used to fuse multi-sensor data to calculate the real-time environmental risk factor E env (t):

[0094]

[0095] Among them, s i (t) is the normalized reading of the i-th sensor, T is the temperature parameter, λ i is the normalization coefficient.

[0096] The technical concept of the present invention is as follows:

[0097] Specifically, historical accident data, flight logs, and real-time data of the operating environment are analyzed, and potential risk factors in drone operations are identified based on the three dimensions of "man-machine-environment", and a drone operation risk evolution network model is constructed. Taking into account the dynamic changes in risk propagation over time, the importance of risk nodes is comprehensively evaluated by improving the calculation parameters such as node degree, closeness centrality, betweenness centrality, and page ranking (PR) algorithm; at the same time, risk edges are evaluated based on edge betweenness under different environmental risks, and the critical paths of risk propagation in different environments are determined. The present invention introduces a comprehensive evaluation method based on the risk loss coefficient, quantitatively analyzes the impact of node degree, closeness centrality, betweenness centrality, page ranking (PR) algorithm, and edge betweenness on drone operations, and determines key risk nodes and key propagation paths.

[0098] The specific steps are as follows:

[0099] (1) Identification of risk factors

[0100] By analyzing historical accident data, flight logs and real-time data of the operating environment, we identify potential risk factors in the operation of drones based on the three dimensions of "man-machine-environment", including human factors, drone's own performance risk factors and operating environment risk factors.

[0101] (2) Constructing a risk evolution network model

[0102] The identified potential risk factors are regarded as network nodes, and the transmission relationships between risk factors are regarded as network edges, and a network model of drone operation risk evolution is constructed.

[0103] (3) Model parameter calculation

[0104] Considering the dynamic changes of risk propagation over time, the importance and randomness of risk nodes are comprehensively evaluated by improving the calculation parameters such as node degree, proximity centrality, betweenness centrality and page ranking (PR) algorithm; at the same time, based on the edge betweenness under different environmental risks, risk edges are evaluated to determine the key paths of risk propagation in different environments.

[0105] (IV) Risk accident tracing analysis

[0106] Based on the dynamic changes of risk propagation over time, this paper introduces the time attenuation factor, environmental risk change factor and risk loss coefficient, comprehensively analyzes multiple network characteristic indicators, evaluates the importance of nodes, determines the critical path of risk propagation and extracts the key risk propagation chain.

[0107] Example 1: Risk Factor Identification

[0108] Identify potential risk factors in drone operations from the three dimensions of "man-machine-environment", as follows:

[0109] Human factors include inspection and maintenance errors, command center scheduling errors, flight dynamic monitoring deviations, dangerous goods loading, illegal interference, aircraft operation errors, operator inexperience, failure to comply with standard operating procedures, and trajectory planning defects.

[0110] The performance factors of the drone itself include electronic instrument failure, flight program failure, fuselage damage, power system failure, communication system failure, navigation and positioning system failure, flight control system failure, power supply system failure, insufficient battery life, battery overheating, aging of mechanical parts, design and manufacturing defects, too many total flight hours of the aircraft, immature aircraft product technology, excessive aircraft load, and aircraft emergency system failure.

[0111] Operating environment factors include bird strikes, strong winds / gusts / wind shear, thunderstorms, showers, icing, magnetic field interference, aerial obstacles, excessively high ambient temperatures, low visibility, terrain obstacles, and flight conflict risks.

[0112] Example 2: Constructing a risk evolution network model

[0113] The above risk factors are regarded as network nodes, and the transmission relationship between risk factors is regarded as network edges. A UAV operation risk evolution network model is constructed, as shown in Figure 2 shown. Figure 2 In the example, each time a risk is transferred, the edge becomes thicker. The thickness of the edge represents the number of times the risk is transferred, that is, the weight of the edge. The thicker the edge, the more risk types are transferred during the risk transfer process.

[0114] The UAV operation risk evolution network model contains 44 nodes and 335 edges, among which the nodes are divided into human factor nodes (A1-A9), UAV performance nodes (B1-B16), operation environment nodes (C1-C8), and 12 risk event nodes (D1-D12), as shown in Table 1.

[0115] Table 1. Risk factors for drone operations

[0116] node Risk factors node Risk factors A1 Check for maintenance errors B15 The aircraft is overloaded A2 Command center dispatch error B16 Aircraft emergency system failure A3 Flight dynamics monitoring deviation C1 Bird strike A4 Dangerous goods loading C2 Strong winds / gusts / wind shear, thunderstorms, showers, icing A5 Unlawful interference C3 Magnetic field interference A6 Aircraft operation errors C4 Aerial obstacles A7 Inexperienced operators C5 Ambient temperature is too high A8 Failure to follow standard operating procedures C6 Low visibility A9 Trajectory planning flaws C7 Terrain obstacles B1 Electronic instrument failure C8 Flight conflict risk B2 Flight procedure failure D1 Aircraft stall B3 Damage to the fuselage D2 Aircraft yaw B4 Power system failure D3 Aircraft loss of control B5 Communication system failure D4 Fire B6 Navigation and positioning system failure D5 explode B7 Flight control system failure D6 mid-air collision B8 Power supply system failure D7 forced landing B9 Insufficient battery life D8 Disintegration in the air B10 Battery overheating D9 Ground collision B11 Aging of mechanical parts D10 crash B12 Design and manufacturing defects D11 Third-party injury B13 The total flight hours of the aircraft are too high D12 Third-party property damage B14 Aircraft product technology is immature

[0117] The ISM explanatory structure model is used to divide risk factors into five levels: fundamental factors (fifth level), structural factors (fourth level), indirect factors (third level), direct factors (second level), and risk events (first level). This allows for a more direct understanding of the relationship between each risk factor and a risk event. For example, from the perspective of risk prevention, the effect of risk prevention can be achieved by specifically disconnecting direct factors from risk events. Figure 3 shown.

[0118] Example 3: Model parameter calculation

[0119] Node degree calculation: The node degree is an important metric. Nodes with higher degrees are considered more important in the network. Considering the fact that risks change over time, the time node degree K is introduced. i (t). Therefore, the degree of node i is expressed as, and the calculation is as follows:

[0120]

[0121] Where K i (t) represents the time node degree of node i; N represents the total number of network nodes. ij (t) indicates whether there is a risk propagation edge between nodes i and j at time t. For example, during the flight of a drone, as time passes, the battery power gradually decreases and the risk of battery failure increases. The risk propagation path related to the battery may change dynamically, which can reflect the dynamic characteristics of the risk propagation network in real time.

[0122] Calculation of proximity centrality: proximity centrality indicates the average distance between a node and other nodes in the network. The larger the value of a node's proximity centrality, the more important the node is. Nodes with high proximity centrality can establish connections with other nodes in the network more quickly, and the risk spreads relatively quickly in these nodes. Introducing the risk transmission intensity r ij , the path with high risk propagation intensity has a greater impact on the node proximity centrality and can more accurately evaluate the importance of the node in risk propagation. The calculation is shown in the following formula:

[0123]

[0124] Where, d ij The number of edges in the shortest path starting from node i and ending at node j; n represents the total number of nodes; r ij Represents the risk transmission intensity from node i to node j.

[0125] Betweenness centrality calculation: Betweenness centrality measures the proportion of all shortest paths in a network that pass through a node. When risks spread, they often pass through these key nodes between different subgroups. The calculation is shown in the following formula:

[0126]

[0127] Where σ(s, t) represents the total number of shortest paths between nodes s and t; σ(s, t|v) represents the number of shortest paths from node s to node t passing through node v.

[0128] PageRank (PR) algorithm calculation: The PageRank (PR) algorithm is used to simulate a random walk on a network graph, with each step randomly jumping from the current node to the next node along a directed edge. By calculating the steady-state probability of this random walk visiting each node over a long period of time, the PR value of each node in the network model can be obtained. To improve simulation accuracy, a dynamic decay factor d is introduced, which represents the probability of a node continuing to traverse backward at any time. The calculation is shown in the following formula:

[0129] d=d0+Δd×T(t);

[0130] T(t)=e -λt ;

[0131]

[0132] Where: PR(p i ) is the PR value of the node, n is the total number of nodes; d0 is the initial attenuation factor, Δd is the adjustment amplitude, usually 0.25, T(t) is the time function, λ is the preset constant, usually 0.1. For example, during the intense flight mission phase, when risk spreads quickly, d can be appropriately reduced to highlight the impact of recent risk spread; M(p i ) is the same as p j connected nodes; w(p i ) is (p i ,p j )’s edge weight; L(p j ) is p j The degree of the node.

[0133] Edge betweenness calculation: Introducing environmental risk factor E into edge betweenness calculationenv (j,k,i), taking into account the impact of environmental factors on risk propagation paths. The degree of impact of specific environmental conditions (such as weather and magnetic fields) on the risk propagation from node j to k through edge i is determined. In severe weather, some risk propagation paths are more affected. This calculation formula can more accurately identify key risk edges. The calculation is shown in the following formula:

[0134] E i =Σ j,k∈V E env (j,k,i)×n jk (i);

[0135] Where E i is the edge betweenness of i; E env (j, k, i) are environmental risk factors; n jk (i) is the number of paths between node j and node k that pass through edge i.

[0136] Example 4: Risk Factor Hierarchy

[0137] Constructing the adjacency matrix: The adjacency matrix is used to describe the relationship between the various influencing factors. Determine the relationship between the influencing factors and construct the adjacency matrix A = (bij)n × n, which is used to express the logical relationship between the factors in the form of a matrix, where n = 44. According to the rules of the adjacency matrix, the relationship between two factors is expressed by the matrix element b ij express:

[0138]

[0139] Reachability matrix calculation: The reachability matrix R represents the connection channel between one factor and another in the system, and is the data interpretation of the hierarchical model. The reachability matrix calculation requires the satisfaction of Boolean operations, where I is the identity matrix:

[0140] (A+I)n+1=(A+I)n≠......(A+I)2≠(A+I);

[0141] Hierarchical calculation: Hierarchical structure division requires factor extraction and hierarchical construction based on the decomposition results. All factors are decomposed and divided into several levels based on the reachability matrix to construct a hierarchical structure of influencing factors.

[0142] In the first step, the set of elements corresponding to the columns where all matrix elements in the i-th row of the reachable matrix R are 1 is defined as the reachable set M;

[0143] The second step is to define the set of elements corresponding to all rows where the matrix elements in the i-th row of the reachable matrix R are 1 as the predecessor set N;

[0144] The third step is to define the common set H of the reachable set and the antecedent set, where H = M ∩ N. When the condition H = N is met, the corresponding hierarchical factors are extracted. Finally, the drone operation risks are divided into five levels: fundamental factors, structural factors, indirect factors, direct factors, and risk events.

[0145] Example 5: Risk Accident Source Analysis

[0146] By comprehensively analyzing multiple network characteristic indicators, we delve into the critical transmission processes between system risk factors. Using five key evaluation metrics—node degree, closeness centrality, betweenness centrality, PageRank (PR) algorithm, and edge betweenness—we identify key risk nodes and edges, extract key risk transmission chains, and trace the root causes of accidents.

[0147] By accurately tracing the root causes of accidents in real time, we can effectively avoid similar drone accidents, achieving the goal of risk prevention and control. For example, a certain type of drone accident is frequently reported. After tracing the source, it is found that the root cause is the substandard manufacturing process of the drone's battery, resulting in frequent crashes and loss of life and property. To address the battery quality risk, we can choose to replace drones with other established brands to complete the mission, request the manufacturer to upgrade the battery, and other measures to prevent such accidents from happening again. In other words, tracing the source is the foundation of prevention, and the two are progressive rather than contradictory.

[0148] Comprehensively analyze each transmission path in the key risk transmission chain to clarify the risk transmission mechanism and influencing factors between different nodes. In order to make the comprehensive assessment focus more on serious risks, the risk loss coefficient is introduced. The losses caused by different risk events are determined, and the risk loss function is shown in the following formula:

[0149]

[0150] Where S i is the risk loss function, K i 、C i 、B i PR(p i ), E i are the time node degree, closeness centrality, betweenness centrality, page ranking value, and edge betweenness of node i, respectively; α, β, γ, δ, and μ are the weight coefficients of time node degree, closeness centrality, betweenness centrality, page ranking value, and edge betweenness, respectively, all in the range of [0,1], and α+β+γ+δ+μ=1; is the risk loss coefficient. α, β, γ, δ, and μ can be determined by the hierarchical analysis method combined with expert experience. The core role of the risk loss function is to quantitatively analyze the comprehensive risk contribution of each node in the UAV operation risk network, so as to realize the identification of key risk nodes and transmission paths and the accurate tracing of risk accidents. That is, the risk loss function is the core hub of the technical chain of the present invention to realize "multi-dimensional risk dynamic modeling → real-time quantitative evaluation → accurate tracing decision-making". Its essence is to transform the complex UAV operation risk system into a calculable and interpretable quantitative indicator through mathematical modeling, which solves the problems of strong subjectivity and poor timeliness of risk assessment in traditional methods, and provides a data-driven scientific basis for the safe operation of UAVs.

[0151] Example 6: Adaptive Adjustment of Dynamic Parameters

[0152] By introducing a dynamic adjustment mechanism, the model automatically adjusts the weights of nodes and edges in the network based on real-time flight data and environmental changes, thereby better adapting to the dynamic changes in drone operations and improving the accuracy and timeliness of risk accident tracing. Furthermore, the model automatically optimizes parameter calculation methods during continuous operation, adjusting algorithm parameters based on actual risk assessment feedback, improving the accuracy and adaptability of parameter calculations and providing a more reliable basis for risk accident tracing.

[0153] (1) Dynamic risk attenuation coefficient prediction based on LSTM

[0154] An LSTM network is used to model the degradation data of key drone components, such as batteries and motors, in a time series fashion to predict the dynamic risk attenuation coefficient at different time points. LSTM networks can capture long-term dependencies in time series data and are suitable for handling dynamic changes in drone operations.

[0155] 1. LSTM Network Structure

[0156] The LSTM network consists of an input gate, a forget gate, and an output gate, and can effectively handle long-term dependencies in time series data. Its basic unit structure is as follows:

[0157]

[0158] Among them, i t 、f t and o t are the activation values of the input gate, forget gate, and output gate respectively; is the value of the candidate memory cell; c t is the memory unit at the current moment; h t is the hidden state at the current moment; W i 、W f 、W o and Wc is the weight matrix; b i 、b f 、b o and b c is the bias term; σ is the Sigmoid activation function; tanh is the hyperbolic tangent activation function; x t is the input feature at the current moment; h t-1 is the hidden state at the previous moment; c t-1 It is the memory unit of the previous moment.

[0159] 2. Dynamic Risk Attenuation Coefficient Prediction Formula

[0160] The dynamic risk attenuation coefficient prediction formula based on the LSTM network is as follows:

[0161] T(t)=σ(W·h t +b);

[0162]

[0163] Where t is the cumulative time of risk propagation (unit: seconds); T(t) is the dynamic risk attenuation coefficient; W is the weight matrix from the LSTM hidden state to the output; b is the bias term; the initial values of the weight matrix W and the bias term b are combined with the physical meaning of drone risk propagation, and the initial weights are set by expert scoring to ensure that the model conforms to the prior logic; σ is the Sigmoid function; h t is the hidden state of LSTM at time t.

[0164] 3. Input features and output labels

[0165] The input feature is the timestamp of historical risk data, and the auxiliary input features are yaw distance, wind speed, battery voltage, temperature, charge and discharge times, ambient temperature, etc. The output is the risk attenuation coefficient T(t+1) at the next moment.

[0166] The attenuation coefficient is defined to describe the characteristic that risk gradually weakens over time. The impact of risk propagation will decay as time t increases. The sigmoid function σ(x) used in this invention has a value range of T(t) (0, 1). By adjusting the value of the bias term b, the risk attenuation coefficient satisfies that when t approaches 0, T(t) ≈ 1 (risk is not attenuated); when t increases, T(t) gradually approaches 0 (risk decays to near extinction).

[0167] For example, a minor fault that occurred early on caused a slight overheating of the battery but did not cause a subsequent accident, and its contribution to the current risk will decrease over time. Assuming b = 10, W = [0.1], then:

[0168] t = 0 seconds, When t = 50 seconds; The risk has basically subsided.

[0169] 1) Data preprocessing

[0170] Step 1: Collect time series data

[0171] Collect the timestamps and auxiliary features (wind speed, yaw distance, power) of historical risk events, such as predicting the attenuation coefficient of strong wind → yaw risk.

[0172] Time t (seconds) Strong wind duration (seconds) Yaw distance (m) 0 0 0 30 30 20 60 60 15 90 90 5

[0173] Step 2: Construct input-label pairs

[0174] Sliding window: Set the window size to n, take n consecutive time step data as input, and the data of the next time step as the label.

[0175] Input features: [T(tn), T(t-n+1), …, T(t)];

[0176] Output label: T(t+1);

[0177] For example, input features: [strong wind duration, yaw distance]; label: the next moment T(t+1). Assume W = [0.5, -0.3] (strong wind duration weight is positive, yaw distance weight is negative), b = 2, the weight matrix W and the bias term b are continuously optimized through model training to make T(t) fit the real attenuation curve, then

[0178] The longer the strong wind lasts, the larger T(t) is, and the slower the risk decays; the smaller the yaw distance is, the smaller T(t) is, and the risk has decayed.

[0179] Step 3: Data Normalization

[0180] The timestamps of historical risk events and auxiliary features (yaw distance, wind speed, battery power, etc.) are normalized and scaled to the range of [0, 1] to avoid slow model convergence due to differences in feature value ranges and improve training efficiency.

[0181] (2) Real-time prediction of environmental risk factors based on reinforcement learning

[0182] A reinforcement learning (RL) model was constructed to dynamically adjust edge betweenness weights based on real-time environmental data (wind speed, magnetic field strength, and precipitation level). The RL model can dynamically adjust its strategy based on environmental feedback to adapt to changing environmental conditions.

[0183] 1. Reinforcement Learning Model Architecture

[0184] It adopts an Actor-Critic architecture based on PPO (Proximal Policy Optimization), in which the Actor network outputs actions and the Critic network evaluates the state value.

[0185] State space: current environmental parameters (wind speed, magnetic field strength, precipitation level) + drone flight status (altitude, speed).

[0186] Action space: The adjustment increment Δw for adjusting the edge betweenness weight is limited to the range of [-0.1, 0.1] to avoid sudden changes.

[0187] Reward function: Accuracy of risk propagation path prediction (verified by historical accident data).

[0188] 2. Reinforcement Learning Update Formula

[0189] The update formula of the reinforcement learning model is as follows:

[0190]

[0191] V(s t ) update = V(s t )+η·(R t +γV(s t+1 )-V(s t ));

[0192] Where θ is the parameter of the Actor network; η is the learning rate, which is adjusted according to the convergence speed and is set to 0.01. is the gradient of the parameter θ; π θ (a t |s t ) is in state s t Next take action a t probability; is the estimated value of the advantage function; V(s t ) is the critic network's response to state s t The value estimate of R t is in state s t Next take action a t The reward obtained after γ is the discount factor; s t+1 Is to perform action a t Then transfer to the new state.

[0193] 3. Real-time prediction formula for environmental risk factors

[0194] The real-time prediction formula for environmental risk factors based on the reinforcement learning model is as follows:

[0195] State space: s=[v,B,r,h,vd ,Esoc];

[0196] Action space: a t =Δw;

[0197] Reward function: R t =R accuracy +R safety -R penalty ;

[0198] Among them, s is the state space dimension, a t is the action space dimension, R t is the reward function, v is the wind speed; B is the magnetic field strength; r is the precipitation level; h is the drone height; v d is the speed of the drone; Esoc is the remaining battery power; Δw is the adjustment increment of the edge betweenness weight; R accuracy is the accuracy reward; R safety is a safety reward; R penalty is a stability penalty.

[0199] (3) Parameter adaptive adjustment based on feedback optimization

[0200] The model automatically optimizes the parameter calculation method during continuous operation, adjusts the algorithm parameters based on actual risk assessment feedback, and improves the accuracy and adaptability of parameter calculation.

[0201] 1. Feedback optimization mechanism

[0202] The core of the feedback optimization mechanism is to compare the risk assessment results with the actual accident data, calculate the assessment error, and dynamically adjust the model parameters based on the error.

[0203] 2. Parameter adaptive adjustment

[0204] The parameter adaptive adjustment formula is as follows:

[0205] R error =R predicted +R actual ;

[0206] θ new =θ old +ηR error ;

[0207] Among them, R error is the evaluation error, R predicted is the risk value predicted by the model; R actual is the actual risk value; θ new are the old model parameters; θ old is the updated model parameter; η is the learning rate, which is adjusted according to the convergence speed and is set to 0.01.

[0208] 3. Dynamically adjust the weights of nodes and edges

[0209] According to real-time flight data and environmental changes, the weights of nodes and edges in the network are dynamically adjusted, so that the model can better adapt to the dynamic changes during the operation of the UAV.

[0210]

[0211] in, is the old weight of node i; is the new weight of node i; f(t) is the dynamic risk attenuation coefficient; β 节点 is the node weight adjustment coefficient; is the old weight of the edge; is the new weight of the edge; Δw is the adjustment increment of the edge betweenness weight; γ 边 is the edge weight adjustment coefficient.

[0212] (4) Technical Effect

[0213] Through this dynamic adjustment mechanism, the model can automatically adjust the weights of nodes and edges in the network based on real-time flight data and environmental changes, improving the accuracy and timeliness of risk accident tracing. At the same time, the model automatically optimizes parameter calculation methods during continuous operation, adjusting algorithm parameters based on actual risk assessment feedback, improving the accuracy and adaptability of parameter calculations. The specific technical effects are as follows:

[0214] 1. From experience-driven to data-driven: The adjustment of dynamic parameters is shifted from experience-driven to data-driven, improving the model's adaptability to different scenarios.

[0215] 2. Reduce manual intervention: Reduce manual intervention and respond to sudden environmental changes (such as strong winds) in real time.

[0216] 3. Reduce errors: LSTM+RL dynamic prediction effectively reduces errors.

[0217] 4. Improve risk path coverage: Hypergraph modeling breaks through the limitations of traditional single-layer networks, supports multi-hop risk propagation analysis, and improves risk path coverage.

[0218] 5. Reduce subjective bias: Eliminate subjective bias in manually defined nodes / edges and improve model objectivity.

[0219] 6. Improved risk perception accuracy: Quantifying environmental risk factors with objective standards reduces reliance on subjective experience. Multi-sensor data fusion improves risk perception accuracy in complex environments.

[0220] Example 7: Multi-layer network modeling and cross-dimensional risk propagation

[0221] Hierarchical network structure design:

[0222] Human factors layer: nodes are operator behaviors (false touches, command delays) and ground station status; edges are the impact of operational errors on the propagation of command links.

[0223] Airborne layer: Nodes are drone hardware (batteries, motors, sensors) and software modules; edges are fault transmission relationships (such as insufficient battery voltage causing the motor to stop).

[0224] Environmental layer: Nodes are weather (wind speed, precipitation), electromagnetic interference sources, and geographical obstacles; edges are the paths by which environmental factors affect drones.

[0225] Cross-layer interaction modeling:

[0226] Use hypergraphs to represent cross-layer risk propagation, for example:

[0227] Hyperedge 1: Operator’s accidental touch (human factor layer) → command delay (airborne layer) → obstacle avoidance failure (environmental layer).

[0228] Define cross-layer weight W cross =εE env +ξT(t);

[0229] ε+ξ=1, optimized by gradient descent method, with initial values of ε=0.6, ξ=0.4 (based on historical accident data statistics).

[0230] Strengthen the dynamic learning parameters ε and ξ, which are cross-layer interaction weight coefficients and are updated through Q-learning:

[0231] α t+1 =α t +η(R 实际 -R 预测 )·E env ;

[0232] η is the learning rate, which is adjusted according to the convergence speed and is set to 0.01. 实际 This is the immediate reward calculated based on real-world feedback or historical data, reflecting the effect of the current action in the real world. The calculation method is:

[0233]

[0234] Example:

[0235] When the drone triggers the obstacle avoidance action under the magnetic field strength B = 80μT and successfully avoids the obstacle, then R 实际 =0 (no event but risk exists).

[0236] If obstacle avoidance is not triggered and a collision occurs, R 实际 =1.

[0237] R 预测 It is the estimated value of the model's expected reward for the current action, output by the Q-value function in Q-learning or the Critic network, and is calculated as: R 预测 =Q(s t ,a t ) or V(s t );

[0238] Among them, Q(s t ,a t ) is in state s t Next, perform action a t The expected Q value of V(s t ) is state s t The value function (Critic network output).

[0239] Example:

[0240] If the model predicts that the Q value of triggering obstacle avoidance under the magnetic field strength B = 80μT is Q(s t ,a t )=0.8, then R predicted =0.8.

[0241] Technical Effect: Accurately identifies cross-dimensional compound risks (e.g., the crash path caused by the combination of operational error, battery aging, and strong winds). Hypergraph modeling transcends the limitations of traditional single-layer networks and supports multi-hop risk propagation analysis. Hypergraph cross-layer modeling improves risk path coverage.

[0242] Example 8: Data-driven network structure optimization

[0243] Causal network construction:

[0244] Based on historical accident data, the PC algorithm (Peter-Clark) is used to automatically derive the causal relationship between nodes from the data, replacing the 335 manually defined edges.

[0245] Node relationship correction rule: If P(A|B) / P(A)>2.0, then it is considered that there is a causal edge B→A. 2.0 is an empirical value and can be adjusted through statistical significance tests (such as p<0.05).

[0246] Dynamic topology update mechanism:

[0247] Introducing Graph Neural Network (GNN) to update the network structure in real time:

[0248] Node embedding:

[0249] is the embedding vector of node v in the k+1 layer, is the embedding vector of node v in the kth layer, N(v) is the neighbor set of node v, W is the trainable weight matrix, is the embedding vector of node u in the kth layer; CONCAT is the vector concatenation operation;

[0250] Edge weight update: W uv =MLP(h u ,h v );h u is the embedding vector of node u, h v is the embedding vector of node v, MLP is a multi-layer perceptron;

[0251] Recalculate the network topology based on sensor data every 30 seconds;

[0252] Graph Neural Network (GNN) parameters:

[0253] Node embedding dimension: l = 128, which represents the latent space dimension of node features. The larger the node embedding dimension l is, the richer the node feature information that the model can capture;

[0254] Weight matrix W∈Rl×l: learned through end-to-end training;

[0255] Update cycle: 30 seconds, set according to the drone sensor sampling frequency (≥2Hz).

[0256] Technical Effect: Eliminates subjective bias in manually defined nodes and edges, improving model objectivity. Dynamic topology adapts to changes in drone status (e.g., disconnecting non-critical nodes when the battery runs out). PC algorithm combined with GNN dynamic topology reduces subjective bias.

[0257] Example 9: Quantitative Improvement of Environmental Risk Factors

[0258] Environmental risk database construction:

[0259] Establish an environmental parameter-risk weight mapping table:

[0260] Environmental parameters Risk Level <![CDATA[Weight E env > Wind speed>10m / s high 0.9 Magnetic field interference>50μT middle 0.6 Visibility <500m Low 0.3

[0261] Definition: The impact of environmental parameters on risk transmission paths, mapped by table lookup method:

[0262] Environmental parameters Quantification rules (example) Wind speed v <![CDATA[E env =min(0.2v,1.0)]]> Magnetic field strength B <![CDATA[E env =0.01(B-50) + (x + =max(x,0))]]> Visibility <![CDATA[E env =1-z / 1500(z≤1500m)]]>

[0263] Sensor fusion correction:

[0264] Use Kalman filtering to fuse multi-sensor data (barometer, magnetometer, GPS) to calculate environmental risk factors in real time:

[0265] Temperature parameter T: controls the smoothness of weight distribution. The default value is T = 1.0. High temperature (T>1) enhances the influence of small weights, and low temperature (T<1) highlights the dominant sensor.

[0266] Technical Effect: Quantifying environmental risk factors provides objective standards, reducing reliance on subjective experience. Multi-sensor data fusion improves risk perception accuracy in complex environments.

[0267] Example 10: Scenario: A drone experiences a sudden battery voltage drop while performing an inspection in a strong magnetic field. Dynamic parameter adjustment: LSTM detects an abnormal voltage drop rate and triggers the RL model to adjust the environmental weight E. env (Magnetic field risk factor rises to 0.8).

[0268] Multi-layer network warning: Hypergraph identification of cross-layer paths: battery failure (airborne layer) → navigation signal loss (environmental layer) → tower collision risk (human factors layer).

[0269] Data-driven decision-making: GNN updates the network topology, cuts off power to non-essential modules, and prioritizes the operation of the obstacle avoidance system.

[0270] Example 11: Dynamic Parameter Adaptive Adjustment

[0271] Data preparation: Collect historical data on drone battery voltage, temperature, and ambient wind speed, with a time window of n = 10;

[0272] LSTM training: Construct a two-layer LSTM network (hidden layer dimension l = 64), input sequence length tn:t, and output T(t);

[0273] Real-time prediction: Deploy the model to the drone edge computing module and update T(t) every 5 seconds;

[0274] Effect verification: In the battery aging scenario, the risk warning time of the optimized model is 120 seconds earlier than the traditional method.

[0275] Example 12: Cross-layer hypergraph modeling

[0276] Layered definition:

[0277] Human factors nodes: operator command delay, interface mis-touch;

[0278] Airborne layer nodes: battery voltage, motor speed, GPS module;

[0279] Environmental layer nodes: wind speed, magnetic field strength, visibility;

[0280] Hyperedge construction: Define compound risk paths, e.g.

[0281] Hyperedge 1 (operation delay → motor speed drop → strong wind interference → crash risk), initial weight Wcross =0.7;

[0282] Dynamic update: When the magnetic field strength exceeds 50μT, the relevant hyperedge weight is automatically increased to W cross =0.9.

[0283] Example 13: Quantification of environmental factors

[0284] Sensor calibration: The magnetometer range is set to 0-100μT, normalized s i B / 100;

[0285] Weight calculation: In a strong magnetic field environment (B = 80μT), E env =0.01×(80-50)=0.3;

[0286] Risk fusion: If there is a wind speed of 15m / s (E env =0.9), the total environmental risk weight is 0.3+0.9=1.2 (emergency obstacle avoidance is triggered when the threshold value 1.0 is exceeded).

[0287] Example 14: T(t) is the LSTM prediction value

[0288] Input data:

[0289] Battery voltage sequence within time window tn:t: 3.7V, 3.5V, 3.3V, 3.1V (n=4);

[0290] Current environmental risk factor E env =0.8 (wind speed 15m / s).

[0291] LSTM predicts T(t):

[0292] LSTM outputs hidden state ht = [0.6, -0.2, 0.4]; the Sigmoid function is defined as:

[0293] calculate

[0294] Among them, 0.5·h t +0.1=[0.3,-0.1,0.2]+0.1=[0.4,0.0,0.3];

[0295]

[0296]

[0297] Therefore, T(t)≈[0.60,0.50,0.57].

[0298] Calculate cross-layer weights: W cross=0.6×E env +0.4×T(t)=[0.72,0.68,0.708]

[0299] If W cross When >0.7, a high-risk warning is triggered.

[0300] Example 15: Environmental Risk Factor E env The specific process of reinforcement learning (RL) modeling

[0301] 1. Reinforcement Learning Model Selection and Architecture:

[0302] Model type: Actor-Critic architecture based on PPO (Proximal Policy Optimization);

[0303] Selection basis: PPO performs stably in continuous action space and high-dimensional state space and is suitable for dynamic adjustment of Eenv.

[0304] Network structure:

[0305]

[0306] State space: s t =[Env t ,UAV t ].

[0307] Environment parameter Env t : wind speed v∈[0,30]m / s, magnetic field strength B∈[0,100]μT, precipitation level r∈{0(no rain),1(light rain),2(heavy rain)}.

[0308] UAV status t : UAV height h∈[0,500]m, UAV speed v d ∈[0,20]m / s, remaining battery capacity E SOC ∈[0,100]%.

[0309] Action space: a t =ΔE env ∈[-0.1,0.1].

[0310] Definition: Each step adjusts E env The increment is limited to ±0.1 to avoid sudden changes.

[0311] Update rule: E env (t+1)=Clip(E env (t)+a t ,0,1).

[0312] Reward function: Rt =R accuracy +R safety -R penalty ;

[0313] Accuracy Bonus:

[0314] Model prediction value, The true value is inferred based on historical accidents.

[0315] Safety Reward R safe :

[0316]

[0317] Stability penalty: R penalty =0.1·|a t ∣; (suppress frequent adjustments)

[0318] 2. Training Process

[0319] Data preparation:

[0320] Historical data set: Contains time series data of environmental parameters, drone status, and accident records (sampling frequency 1Hz).

[0321] Synthetic data enhancement: Extend extreme scenarios (such as magnetic field surge + strong wind) through Gaussian noise injection and adversarial sample generation.

[0322] Simulation environment construction:

[0323] Use the Gazebo or AirSim simulation platform to simulate the flight scenarios of drones in dynamic environments.

[0324] Risk event trigger: When When ,faults are randomly injected (e.g. GPS signal loss).

[0325] Training steps:

[0326] Initialization: The weights of the Actor and Critic networks are randomly initialized, and the experience replay pool capacity is N = 106.

[0327] Interactive sampling: execute the current policy in the simulation environment and collect trajectories (s t ,a t ,R t ,s t+1 ).

[0328] Strategy Optimization:

[0329] Critic update: Minimize the temporal difference error (TD Error):

[0330] V(s t ) is the Critic network state s t The value estimate of ,γ is the discount factor, s t+1 To perform action a t The new state transferred to later, E is the expected operation;

[0331] Actor update: maximize the PPO objective function:

[0332]

[0333] The advantage function A t =R t +γV(s t+1 )-V(s t ),π θ (a t |s t ) is the Actor network parameter θ, in state s t Next select action a t The probability density of θold is the old strategy network parameter (Actor network before update), ∈ is the clipping threshold, ∈ = 0.2, and Clip is the clipping function.

[0334] Iterative convergence: training until the critic loss L critic <0.01 and the reward curve is stable.

[0335] 3. Implementation Method (Deployment and Verification)

[0336] Model lightweighting: Convert the trained PPO model to the TensorRT engine and deploy it to the drone edge computing unit (such as NVIDIA Jetson).

[0337] Inference latency requirement: single-step prediction < 10ms.

[0338] Real-time prediction process:

[0339] Data acquisition: read data from sensors t (wind speed, magnetic field, battery level, etc.).

[0340] Action generation: Actor network outputs a t , update E env (t+1).

[0341] Risk Decision: If E env >0.7, trigger obstacle avoidance or return command.

[0342] Effect verification:

[0343] Test scenario: strong magnetic field (B = 80μT) superimposed on battery aging (E bat =20%).

[0344] Results comparison table:

[0345] method Risk warning accuracy False positive rate Before RL dynamic adjustment 65％ 25％ After RL dynamic adjustment 92％ 8％

[0346] The above embodiments are intended to illustrate the present invention, not to limit the present invention. Therefore, changes in illustrative values or substitutions of equivalent components should still fall within the scope of the present invention.

[0347] From the above detailed description, it will be clear to those skilled in the art that the present invention can indeed achieve the aforementioned objectives and is in compliance with the provisions of the Patent Law.

[0348] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they become aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as covering the preferred embodiments and all changes and modifications that fall within the scope of the invention. The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

[0349] It should be noted that the above description of the relevant processes is for illustration and explanation only and does not limit the scope of application of this specification. For those skilled in the art, various modifications and changes can be made to the processes under the guidance of this specification. However, such modifications and changes are still within the scope of this specification.

[0350] The basic concepts have been described above. It will be apparent to those skilled in the art after reading this application that the above disclosures are merely illustrative and do not constitute limitations on this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and amendments to this application. Such modifications, improvements, and amendments are suggested in this application and remain within the spirit and scope of the exemplary embodiments of this application.

[0351] At the same time, this application uses specific terms to describe the embodiments of this application. For example, "one embodiment," "an embodiment," and / or "some embodiments" refer to a certain feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "one embodiment," "an embodiment," or "an alternative embodiment" mentioned twice or more in different places in this specification does not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application may be appropriately combined.

[0352] In addition, it will be understood by those skilled in the art that various aspects of the present application can be illustrated and described by a number of patentable categories or situations, including any new and useful combination of processes, machines, products or substances, or any new and useful improvements thereto. Therefore, various aspects of the present application can be implemented entirely by hardware, entirely by software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. The above hardware or software can all be referred to as "units", "modules" or "systems". In addition, various aspects of the present application can take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0353] The computer program code required for the operation of each part of the application can be written in any one or more programming languages, including object-oriented programming languages such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python, conventional procedural programming languages such as C programming language, VisualBasic, Fortran2103, Perl, COBOL2102, PHP, ABAP, dynamic programming languages such as Python, Ruby and Groovy or other programming languages. The program code can be run completely on the user's computer, or run on the user's computer as an independent software package, or run partly on the user's computer and partly on a remote computer, or run completely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer through any network form, such as a local area network (LAN) or a wide area network (WAN), or be connected to an external computer (such as by the Internet), or in a cloud computing environment, or used as a service such as software as a service (SaaS).

[0354] In addition, unless expressly stated in the claims, the order of the processing elements and sequences described in this application, the use of alphanumeric characters, or the use of other names are not intended to limit the order of the processes and methods of this application. Although the above disclosure discusses some embodiments of the invention that are currently considered useful through various examples, it should be understood that such details are only for illustrative purposes, and the attached claims are not limited to the disclosed embodiments. On the contrary, the claims are intended to cover all modifications and equivalent combinations that are consistent with the essence and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a pure software solution, for example, installation on an existing server or mobile device.

[0355] Similarly, it should be noted that in order to simplify the presentation of this disclosure and thereby facilitate understanding of one or more of the invention's embodiments, the foregoing descriptions of the embodiments of this disclosure sometimes combine multiple features into a single embodiment, figure, or description thereof. However, this approach should not be interpreted as reflecting an intention that the claimed subject matter requires more features than expressly recited in each claim. Rather, the subject matter of the invention may possess fewer features than the single embodiment described above.

Claims

1. A multi-dimensional UAV risk accident tracing method based on a dynamic network model, characterized by: include: Risk Factor Identification: Analyze historical accident data, flight logs, and real-time operating environment data to identify potential risk factors in drone operations based on the three dimensions of human-machine-environment. Potential risk factors include human factors, drone performance risk factors, and operating environment risk factors. Constructing a risk evolution network model: Treating potential risk factors as network nodes and the transmission relationships between risk factors as network edges, constructing a UAV operation risk evolution network model; Model parameter calculation: Based on the UAV operation risk evolution network model, the time node degree theory is introduced to calculate the node degree, the risk propagation intensity is introduced to calculate the closeness centrality, the betweenness centrality is calculated, the dynamic attenuation factor is introduced to calculate the webpage ranking value, and the environmental risk factor is introduced to calculate the edge betweenness; Risk accident source tracing analysis: Based on the dynamic changes in risk propagation over time, the risk loss coefficient is introduced, and a variety of network characteristic indicators are comprehensively analyzed to evaluate the importance of nodes, determine the critical paths of risk propagation, and extract key risk propagation chains to obtain analysis results; Introducing the risk loss coefficient, the formula for comprehensively analyzing multiple network characteristic indicators is as follows: Where S i is the risk loss function, K i 、C i 、B i PR(p i ), E i are the time node degree, closeness centrality, betweenness centrality, page ranking value, and edge betweenness of node i, respectively; α, β, γ, δ, and μ are the weight coefficients of time node degree, closeness centrality, betweenness centrality, page ranking value, and edge betweenness, respectively, all in the range of [0,1], and α+β+γ+δ+μ=1; is the risk loss coefficient.

2. The multi-dimensional risk accident tracing method for drones based on a dynamic network model according to claim 1 is characterized in that: The formula for calculating node degree by introducing the time node degree theory is as follows: Where K i (t) represents the time node degree of node i; N represents the total number of network nodes; a ij (t) indicates whether there is a risk propagation edge between nodes i and j at time t. If yes, it takes the value 1; if no, it takes the value 0.

3. The multi-dimensional risk accident tracing method of UAV based on dynamic network model according to claim 1 is characterized by introducing Risk transmission intensity calculation close to centrality C i The formula is as follows: Where, d ij The number of edges in the shortest path starting from node i and ending at node j; n represents the total number of nodes; r ij represents the risk transmission intensity from node i to node j; f ij is the number of times the path from node i to j appears in the historical accident data.

4. The multi-dimensional risk accident tracing method for drones based on a dynamic network model according to claim 1 is characterized in that: Calculate betweenness centrality B i The formula is as follows: Where σ(s, t) represents the total number of shortest paths between nodes s and t; σ(s, t|v) represents the number of shortest paths from node s to node t passing through node v.

5. The multi-dimensional risk accident tracing method for drones based on a dynamic network model according to claim 1 is characterized in that: The formula for calculating the web page ranking value by introducing the dynamic decay factor d is as follows: d=d0+Δd×T(t); T(t)=e -λt ; Where: PR(p i ) is the webpage ranking value of the node, n is the total number of nodes; d0 is the initial attenuation factor, Δd is the adjustment amplitude, T(t) is the time function, λ is the preset constant; M(p i ) is the same as p j connected nodes; w(p i ) is (p i ,p j )’s edge weight; L(p j ) is p j The degree of the node; w(p i ) depends on the risk propagation probability P(i→j) of the edge, that is, the node p i After the occurrence, node p j The conditional probability of this happening is: Among them, N(i∩j) is the node p i and p j The number of simultaneous occurrences, N(i) is the number of nodes p i The number of separate occurrences.

6. The multi-dimensional risk accident tracing method for drones based on a dynamic network model according to claim 1 is characterized in that: The formula for calculating edge betweenness by introducing environmental risk factors is as follows: E i =∑ j,k∈V E env (j,k,i)×n jk (i); Where E i is the edge betweenness of i; E env (j,k,i) are the environmental risk factors of j, k and i; n jk (i) is the number of paths between node j and node k that pass through edge i.

7. The multi-dimensional risk accident tracing method for drones based on a dynamic network model according to claim 5 is characterized in that: Use LSTM network to predict time function T(t); T(t)=σ(W·h t +b); Its hidden state h t The update formula is: h t =LSTM(x t-n:t ); Where σ is the Sigmoid function, W is the weight matrix from the LSTM hidden state to the output, b is the bias term, and x t-n:t is the input feature within the time window.

8. The multi-dimensional risk accident tracing method for drones based on a dynamic network model according to claim 1 is characterized in that: Constructing a risk evolution network model includes: Establish a hierarchical network structure, including the human factors layer, the airborne layer, and the environmental layer; Use hypergraph to represent cross-layer risk propagation path, hyperedge weight W cross Calculated by the following formula: W cross =εE env +ξT(t); ε, ξ are dynamic learning parameters, E env is the environmental risk factor; T(t) is the time function.

9. The multi-dimensional risk accident tracing method for drones based on a dynamic network model according to claim 1 is characterized in that: Use Kalman filtering to fuse multi-sensor data and calculate the real-time environmental risk factor E env (t): Among them, s i (t) is the normalized reading of the i-th sensor, T is the temperature parameter, λ i is the normalization coefficient.

Images