An online prediction method for autonomous landing risk of fixed-wing unmanned aerial vehicle under extreme weather based on streaming learning

Abstract

The application faces fixed-wing unmanned aerial vehicles, and proposes an online prediction method for autonomous landing risk of fixed-wing unmanned aerial vehicles in extreme weather based on streaming learning. It contains four steps: (1) data preprocessing is performed on the collected airborne flight data; (2) a comprehensive landing risk model of the fixed-wing unmanned aerial vehicle is constructed; (3) a landing timing process model of the fixed-wing unmanned aerial vehicle is constructed; and (4) online risk prediction of autonomous landing risk of the fixed-wing unmanned aerial vehicle based on streaming learning.

Description

Technical Field

[0001] This invention addresses the needs of fixed-wing unmanned aerial vehicles (UAVs) and proposes an online prediction method for autonomous landing risks under extreme weather conditions based on streaming learning. The method focuses on the typical task of fixed-wing UAV landing, collecting onboard flight data and proposing an effective data preprocessing method. Based on the processed data, a comprehensive landing risk model for the fixed-wing UAV is constructed. A bidirectional long short-term memory (LSTM) model is used for temporal modeling of the flight data, and a Transformer is combined to perform global modeling of the data, training a basic model for predicting landing risks. Finally, streaming learning is used to update the basic model in real time through eligibility traces and dynamic learning rates, and combined with a risk assessment model, to achieve real-time online prediction of landing risks, providing a basis for safe landing decisions and control of fixed-wing UAVs. This invention belongs to the fields of artificial intelligence and safety engineering. Background Technology

[0002] In recent years, with the continuous development of artificial intelligence technology and the widespread application of fixed-wing UAVs in aerial surveying, logistics delivery, and emergency rescue, the safety of their missions has received increasing attention. The landing phase has the highest accident rate, making real-time risk prediction crucial. However, sudden changes in wind speed and turbulent airflow under extreme weather conditions significantly increase landing risks. Therefore, a real-time online risk assessment method is needed for the landing process of fixed-wing UAVs under extreme weather conditions to accurately predict landing risks at various moments and provide decision support for the safe landing of fixed-wing UAVs.

[0003] Fixed-wing UAVs may encounter risks such as hard landing, runway deviation, and tail strike during landing. Current research on fixed-wing UAV landing risks often analyzes single risks, such as hard landing, failing to accurately quantify the comprehensive risks involved in the landing process. Most methods build risk prediction models under ideal weather conditions, which are unsuitable for extreme weather. Furthermore, existing methods rely on pre-trained models, leading to a significant drop in accuracy and increased bias in extreme weather conditions, making it difficult to accurately quantify landing risks under these conditions. Therefore, this invention proposes an online prediction method for autonomous landing risks of fixed-wing UAVs under extreme weather conditions based on streaming learning. First, onboard data from the entire flight process of the fixed-wing UAV is collected and preprocessed to extract key flight parameters and data for the landing phase. Based on the processed data, a comprehensive landing risk assessment model is constructed. Second, a bidirectional long short-term memory (Bi-LSTM) network is used to perform local temporal modeling of the time-series flight data, and a Transformer structure is used to supplement global information modeling capabilities, training a basic model for landing risk prediction. Finally, by combining streaming learning with qualification traces and dynamic learning rates, the efficiency of parameter updates and the learning process of the stable model are improved. The risk prediction model is updated online in real time, and combined with the risk model, the landing risk of fixed-wing UAVs under extreme weather conditions can be predicted in real time and accurately, providing a basis for the autonomous and safe landing decision and control of fixed-wing UAVs. Summary of the Invention

[0004] This invention provides an online prediction method for the autonomous landing risk of fixed-wing UAVs under extreme weather conditions based on streaming learning, such as... Figure 1 As shown. The main steps of this invention are as follows:

[0005] Step 1: Preprocess the collected airborne flight data;

[0006] The collected airborne flight data is recorded as follows: ,in Indicates the first Each flight data point contains [data points]. A flight parameter, denoted as

[0007] Each flight parameter records data from various moments during the flight.

[0008] Step 1: Data filtering;

[0009] Check the continuity and integrity of flight data, and perform data filtering:

[0010] (1) If the continuous missing time in the time series data is as long as 30 seconds or the total missing time of the data is greater than 1 minute, the data shall be removed;

[0011] (2) Check the flight phase indication parameters in the data. By combining the value of this parameter, we can examine the integrity of the flight phase. This indicates the landing phase, when the data is missing. If the data is invalid, then the data should be removed.

[0012] Step 2: Data alignment and truncation;

[0013] (1) Data alignment is performed using the data with the highest sampling frequency in the flight data as the target; for continuous data, linear interpolation is used to achieve data alignment; for discrete data, forward padding is used to achieve data alignment.

[0014] (2) Combining the landing point indication parameters in each flight parameter, it is denoted as: , obtain the landing point; when The time indicated that the fixed-wing drone had not landed. First time from Mutation At this time, it indicates that the fixed-wing UAV has landed, and this time is recorded as _____. Based on this landing time, data from the landing phase 10 minutes before landing is extracted. The data for each parameter can then be recorded as follows: , ;

[0015] Step 3: Key Parameter Selection

[0016] The maximum correlation minimum redundancy (mRMR) method is used to select flight parameters related to landing risk, and its calculation formula is as follows:

[0017]

[0018] in, For the set of all candidate flight parameters; This represents the number of features to be selected. Indicates the selected feature subset. ; These indicate risk indicators, including the risk of a hard landing, the risk of runway deviation, and the risk of tail-scrape. Representing a subset With category Average mutual information, Representing a subset The average value of pairwise mutual information within the entity;

[0019] Step 4: Data Clustering

[0020] Combining the longitude and latitude flight parameters of the landing time from the airborne data, they are respectively denoted as... , Clustering was performed using the density-based noisy applied spatial clustering (DBSCAN) method to obtain flight data from the same airport.

[0021] Step 2: Construct a comprehensive landing risk model for fixed-wing UAVs;

[0022] Step 1: Calculate the mean and variance of the three parameters in the flight parameters: offset, pitch angle, and vertical landing speed, and define the risk level.

[0023] For the processed data, calculate the offset separately. Pitch angle Vertical landing speed The distribution of the three parameters at each time step is denoted as follows: , , Calculate the mean value for both positive and negative values ​​of each parameter. and variance The greater the deviation of the parameter from 0, the greater the risk. The risk level is defined as follows:

[0024] (1) When the absolute value of the parameter is less than At that time, it is considered low risk;

[0025] (2) The absolute value of the parameter is and If the risk level is between these thresholds, it is considered a medium risk.

[0026] (3) When the absolute value of the parameter is greater than At that time, it is considered a high risk.

[0027] Step 2: Calculate the risks of runway drift, tail strike, and hard landing respectively.

[0028] Combining the offset, pitch angle, and vertical landing velocity obtained at various time points, the risk value of each parameter at each time point is obtained through normalization, and denoted as follows: , , The calculation formula is as follows:

[0029]

[0030] in, Indicates the first The parameters are in The corresponding risk value, Indicates the first The parameters are in The maximum value at time 1. Indicates the first The parameters are in The minimum value at time; Indicates the first The parameters are in The value at time;

[0031] Step 3: Construct a comprehensive landing risk model for fixed-wing UAVs

[0032] A comprehensive landing risk model is constructed by combining the three risks of runway deviation, tail strike, and hard landing. Its expression is as follows:

[0033]

[0034]

[0035]

[0036] in, To integrate the landing risk model, This represents the risk function that takes the maximum value. This represents the risk function for calculating the mean.

[0037] Step 3: Modeling the landing sequence of the fixed-wing UAV;

[0038] Step 1: Construct a local temporal model based on a bidirectional long short-term memory network

[0039] The processed landing phase data is input into a bidirectional long short-term memory network, and regularization techniques are introduced to prevent overfitting. The calculation process can be expressed as follows:

[0040]

[0041]

[0042]

[0043] in, This indicates that the feedforward long short-term memory network is in time. The output at that location, This represents the Long Short-Term Memory (LSTM) network function. This indicates regularization techniques. express The probability, This indicates that the inverse long short-term memory network is in time. The output at that location, Indicates vector concatenation. This represents the hidden vector obtained through time-series modeling.

[0044] Step 2: Construct a global time series model based on Transformer

[0045] In the Transformer, the output of the bidirectional Long Short-Term Memory network is transformed based on a learned linear transformation. The vectors are transformed into query vector Q, key vector K, and value vector V, which are used to calculate the similarity between each time step in the time series. , , The calculation formula is as follows:

[0046]

[0047]

[0048]

[0049] in, , , These are the vectors in the model. , , The corresponding weight matrix, , , This is a bias term.

[0050] The dot product similarity between the query vector Q and the key vector K is calculated and normalized using the softmax function to obtain the attention weights at each time step. These attention weights are then applied to the value vector V using the Attention function to obtain the final self-attention output of the Transformer, calculated as follows:

[0051]

[0052] In the formula, This represents the weighted result of self-attention. This indicates the similarity between the query vector and the key vector. This represents the dimension of the key vector.

[0053] The output based on the self-attention mechanism, after processing through residual structure, layer normalization, and feedforward neural network, yields the final output of the Transformer. The computation process is as follows:

[0054]

[0055]

[0056]

[0057]

[0058] in, This represents the intermediate vector after layer normalization. Representation layer normalization; and These represent the outputs of the extended layer and the compressed layer in the feedforward network (FFN), respectively. Indicates the size is The weight matrix, from the input dimension Mapping to Expanded Dimension ; Indicates the size is The weight matrix, from the input dimension Mapping to extended dimensions ; and These represent the bias terms for the FFN expansion and compression layers, respectively. Indicates the encoder layer in time The final output yields the pre-trained prediction model.

[0059] Step 4: Online risk prediction of autonomous landing risk for fixed-wing UAVs based on streaming learning;

[0060] Step 1: Calculate the prediction error

[0061] The actual flight data is input into the pre-trained prediction model for forward propagation, and then the prediction error is calculated and backpropagated. When the prediction error reaches the specified threshold ∆, step 2 is executed; otherwise, step 5 is executed.

[0062] Step 2: Qualification Update

[0063] Combining the gradient obtained from backpropagation, a qualification trace is introduced to improve parameter update efficiency and stabilize the model learning process. The qualification trace update process is represented as follows:

[0064]

[0065] In the formula, For qualification, The decay factor controls the decay rate of the historical gradient. As a cumulative factor for eligibility traces, it modulates the memory strength in response to sudden perturbations. This represents the gradient at the current moment.

[0066] Step 3: Dynamically update the learning rate

[0067] When updating the eligibility trace, the learning rate is updated simultaneously to adapt to unstable data conditions caused by extreme environments. The update formula is as follows:

[0068]

[0069] In the formula, for Learning rate at any given moment The basic learning rate, This is the sensitivity coefficient. The magnitude of the state prediction error for a fixed-wing UAV.

[0070] Step 4: Online update of the prediction model

[0071] The prediction model is dynamically updated by combining eligibility traces and learning rate, and the parameters in the model are updated using the following formula:

[0072]

[0073] in, for Parameter values ​​of the time-based risk prediction model.

[0074] Step 5: Online Risk Prediction

[0075] By combining real-time data streams and prediction models, accurate predictions of the landing process status of fixed-wing UAVs in extreme environments are achieved, and online risk prediction is performed based on the comprehensive landing risk model obtained in step two. Attached Figure Description

[0076] Figure 1. An online prediction method for landing risk of fixed-wing UAVs under extreme weather conditions based on streaming learning.

[0077] Figure 2. Schematic diagram of flight data clustering results

[0078] Figure 3. Schematic diagram of three risk distributions based on statistical analysis.

[0079] Figure 4. Temporal modeling based on bidirectional long short-term memory network

[0080] Figure 5 Global modeling based on Transformer

[0081] Figure 6. Schematic diagram of training / validation loss

[0082] Figure 7 Risk prediction based on streaming learning

[0083] Figure 8. Visualization of risk prediction results Specific implementation methods

[0084] Step 1: Preprocess the collected airborne flight data;

[0085] The collected airborne flight data is recorded as follows: ,in Indicates the first Each flight data point contains [data points]. A flight parameter, denoted as

[0086] Each flight parameter records data from various moments during the flight.

[0087] Step 1: Data filtering;

[0088] Check the continuity and integrity of flight data, and perform data filtering:

[0089] (1) If the continuous missing time in the time series data is as long as 30 seconds or the total missing time of the data is greater than 1 minute, the data shall be removed;

[0090] (2) Check the flight phase indication parameters in the data. By combining the value of this parameter, we can examine the integrity of the flight phase. This indicates the landing phase, when the data is missing. If the data is invalid, then the data should be removed.

[0091] Step 2: Data alignment and truncation;

[0092] (3) Data alignment is performed using the data with the highest sampling frequency in the flight data as the target; for continuous data, linear interpolation is used to achieve data alignment; for discrete data, forward padding is used to achieve data alignment.

[0093] (4) Combining the landing point indication parameters in each flight parameter, it is denoted as: , obtain the landing point; when The time indicated that the fixed-wing drone had not landed. First time from Mutation At this time, it indicates that the fixed-wing UAV has landed, and this time is recorded as _____. Based on this landing time, data from the landing phase 10 minutes before landing is extracted. The data for each parameter can then be recorded as follows: , ;

[0094] Step 3: Key Parameter Selection

[0095] The maximum correlation minimum redundancy (mRMR) method is used to select flight parameters related to landing risk, and its calculation formula is as follows:

[0096]

[0097] in, For the set of all candidate flight parameters; This represents the number of features to be selected. Indicates the selected feature subset. ; These indicate risk indicators, including the risk of a hard landing, the risk of runway deviation, and the risk of tail-scrape. Representing a subset With category Average mutual information, Representing a subset The average value of pairwise mutual information within the entity;

[0098] Step 4: Data Clustering

[0099] Combining the longitude and latitude flight parameters of the landing time from the airborne data, they are respectively denoted as... , Clustering was performed using the density-based noisy applied spatial clustering (DBSCAN) method to obtain flight data from the same airport.

[0100] [Example] A total of 5367 flight data entries were collected from an airline. After data filtering, 3823 flight data entries remained. Landing phase data was extracted, and 29 flight parameters were selected as shown in Table 1. Data obtained based on clustering is as follows... Figure 2 As shown, 703 data points from category 1 were selected as subsequent training data.

[0101] Table 1 Selected Flight Parameters

[0102] FPAC Flight path acceleration CAS Calculate airspeed BLAC longitudinal acceleration of the body TAS vacuum speed CTAC Lateral trajectory acceleration PI Dynamic pressure TH True heading PS static pressure MH Magnetic heading PT Full pressure GS Ground speed AOAI Indicator Angle of Attack TRK True track angle AOAC Correcting the angle of attack TRKM Magnetic track angle BAL1 Correcting air pressure altitude WS wind speed GLS Glide slope deviation WD True wind direction AIL_1 Left Auxiliary Wing Position ALT Pressure height ELEV_1 Left elevator position ALTR height change rate FLAP trailing edge flap position ROLL Roll angle PTRM Pitch trim position N1_1 Fan speed 1 LATP Latitude MACH Mach number

[0103] Step 2: Construct a comprehensive landing risk model for fixed-wing UAVs;

[0104] Step 1: Calculate the mean and variance of the three parameters in the flight parameters: offset, pitch angle, and vertical landing speed, and define the risk level.

[0105] For the processed data, calculate the offset separately. Pitch angle Vertical landing speed The distribution of the three parameters at each time step is denoted as follows: , , Calculate the mean value for both positive and negative values ​​of each parameter. and variance The greater the deviation of the parameter from 0, the greater the risk. The risk level is defined as follows:

[0106] (1) When the absolute value of the parameter is less than At that time, it is considered low risk;

[0107] (2) The absolute value of the parameter is and If the risk level is between these thresholds, it is considered a medium risk.

[0108] (3) When the absolute value of the parameter is greater than At that time, it is considered a high risk.

[0109] Step 2: Calculate the risks of runway drift, tail strike, and hard landing respectively.

[0110] Combining the offset, pitch angle, and vertical landing velocity obtained at various time points, the risk value of each parameter at each time point is obtained through normalization, and denoted as follows: , , The calculation formula is as follows:

[0111]

[0112] in, Indicates the first The parameters are in The corresponding risk value, Indicates the first The parameters are in The maximum value at time 1. Indicates the first The parameters are in The minimum value at time; Indicates the first The parameters are in The value at time;

[0113] Step 3: Construct a comprehensive landing risk model for fixed-wing UAVs

[0114] A comprehensive landing risk model is constructed by combining the three risks of runway deviation, tail strike, and hard landing. Its expression is as follows:

[0115]

[0116]

[0117]

[0118] in, To integrate the landing risk model, This represents the risk function that takes the maximum value. This represents the risk function for calculating the mean.

[0119] [Example] Combining the filtered 3823 flight data points, the distribution of the three risks at various times is obtained as follows: Figure 3 As shown.

[0120] Step 3: Modeling the landing sequence of the fixed-wing UAV;

[0121] Step 1: Construct a local temporal model based on a bidirectional long short-term memory network

[0122] The processed landing phase data is input into a bidirectional long short-term memory network, and regularization techniques are introduced to prevent overfitting. The calculation process can be expressed as follows:

[0123]

[0124]

[0125]

[0126] in, This indicates that the feedforward long short-term memory network is in time. The output at that location, This represents the Long Short-Term Memory (LSTM) network function. This indicates regularization techniques. express The probability, This indicates that the inverse long short-term memory network is in time. The output at that location, Indicates vector concatenation. This represents the hidden vector obtained through time-series modeling, such as Figure 4 As shown.

[0127] Step 2: Construct a global time series model based on Transformer

[0128] In the Transformer, the output of the bidirectional Long Short-Term Memory network is transformed based on a learned linear transformation. The vectors are transformed into query vector Q, key vector K, and value vector V, which are used to calculate the similarity between each time step in the time series. , , The calculation formula is as follows:

[0129]

[0130]

[0131]

[0132] in, , , These are the vectors in the model. , , The corresponding weight matrix, , , This is a bias term.

[0133] The dot product similarity between the query vector Q and the key vector K is calculated and normalized using the softmax function to obtain the attention weights at each time step. These attention weights are then applied to the value vector V using the Attention function to obtain the final self-attention output of the Transformer, calculated as follows:

[0134]

[0135] In the formula, This represents the weighted result of self-attention. This indicates the similarity between the query vector and the key vector. This represents the dimension of the key vector.

[0136] The output based on the self-attention mechanism, after processing through residual structure, layer normalization, and feedforward neural network, yields the final output of the Transformer. The computation process is as follows:

[0137]

[0138]

[0139]

[0140]

[0141] in, This represents the intermediate vector after layer normalization. Representation layer normalization; and These represent the outputs of the extended layer and the compressed layer in the feedforward network (FFN), respectively. Indicates the size is The weight matrix, from the input dimension Mapping to Expanded Dimension ; Indicates the size is The weight matrix, from the input dimension Mapping to extended dimensions ; and These represent the bias terms for the FFN expansion and compression layers, respectively. Indicates the encoder layer in time The final output yields the pre-trained prediction model, such as... Figure 5 As shown;

[0142] [Example] Combining 703 flight data points obtained from clustering, a pre-trained prediction model is trained using a bidirectional long short-term memory network and a Transformer network structure. The changes in training / validation loss during the training process are as follows: Figure 6 As shown.

[0143] Step 4: Online risk prediction of fixed-wing UAV landing risk based on streaming learning;

[0144] Step 1: Calculate the prediction error

[0145] The actual flight data is input into the pre-trained prediction model for forward propagation, and then the prediction error is calculated and backpropagated. When the prediction error reaches the specified threshold ∆, step 2 is executed; otherwise, step 5 is executed.

[0146] Step 2: Qualification Update

[0147] Combining the gradient obtained from backpropagation, a qualification trace is introduced to improve parameter update efficiency and stabilize the model learning process. The qualification trace update process is represented as follows:

[0148]

[0149] In the formula, For qualification, The decay factor controls the decay rate of the historical gradient. As a cumulative factor for eligibility traces, it modulates the memory strength in response to sudden perturbations. This represents the gradient at the current moment.

[0150] Step 3: Dynamically update the learning rate

[0151] When updating the eligibility trace, the learning rate is updated simultaneously to adapt to unstable data conditions caused by extreme environments. The update formula is as follows:

[0152]

[0153] In the formula, for Learning rate at any given moment The basic learning rate, This is the sensitivity coefficient. The magnitude of the state prediction error for a fixed-wing UAV.

[0154] Step 4: Online update of the prediction model

[0155] The prediction model is dynamically updated by combining eligibility traces and learning rate, and the parameters in the model are updated using the following formula:

[0156]

[0157] in, for Parameter values ​​of the time-based risk prediction model.

[0158] Step 5: Online Risk Prediction

[0159] By combining real-time data streams and prediction models, accurate predictions of the landing process status of fixed-wing UAVs in extreme environments are achieved. Furthermore, online risk prediction is performed based on the comprehensive landing risk model obtained in step two. Figure 7 As shown.

[0160] [Example] Risks predicted based on streaming learning, such as Figure 8 As shown.

Claims

1. A method for online prediction of autonomous landing risk of fixed-wing UAVs under extreme weather conditions based on streaming learning, characterized in that... It includes the following steps: Step 1: Preprocess the collected airborne flight data; The collected airborne flight data is recorded as follows: ,in Indicates the first Each flight data point contains [data points]. A flight parameter, denoted as Each flight parameter records data from various moments during the flight. Step 1: Data filtering; Check the continuity and integrity of flight data, and perform data filtering: (1) If the continuous missing time in the time series data is as long as 30 seconds or the total missing time of the data is greater than 1 minute, the data shall be removed; (2) Check the flight phase indication parameters in the data. By combining the value of this parameter, we can examine the integrity of the flight phase. This indicates the landing phase, when the data is missing. If the data is invalid, then remove that data; Step 2: Data alignment and truncation; (1) Data alignment is performed using the data with the highest sampling frequency in the flight data as the target; for continuous data, linear interpolation is used to achieve data alignment; for discrete data, forward padding is used to achieve data alignment. (2) Combining the landing point indication parameters in each flight parameter, it is denoted as: , obtain the landing point; when The time indicated that the fixed-wing drone had not landed. First time from Mutation At this time, it indicates that the fixed-wing UAV has landed, and this time is recorded as _____. Based on this landing time, data from the landing phase 10 minutes before landing is extracted. The data for each parameter can then be recorded as follows: , ; Step 3: Key Parameter Selection The maximum correlation minimum redundancy (mRMR) method is used to select flight parameters related to landing risk, and its calculation formula is as follows: in, For the set of all candidate flight parameters; The number of candidate features; Indicates the selected feature subset. ; These indicate risk indicators, including the risk of a hard landing, the risk of runway deviation, and the risk of tail-scrape. Representing a subset With category Average mutual information, Representing a subset The average value of pairwise mutual information within the entity; Step 4: Data Clustering Combining the longitude and latitude flight parameters of the landing time from the airborne data, they are respectively denoted as... , Clustering was performed using the density-based noisy applied spatial clustering (DBSCAN) method to obtain flight data from the same airport; Step 2: Construct a comprehensive landing risk model for fixed-wing UAVs; Step 1: Calculate the mean and variance of the three parameters in the flight parameters: offset, pitch angle, and vertical landing speed, and define the risk level. For the processed data, calculate the offset separately. Pitch angle Vertical landing speed The distribution of the three parameters at each time step is denoted as follows: , , Calculate the mean value for both positive and negative values ​​of each parameter. and variance The greater the deviation of the parameter from 0, the greater the risk. The risk level is defined as follows: (1) When the absolute value of the parameter is less than At that time, it is considered low risk; (2) The absolute value of the parameter is and If the risk level is between these thresholds, it is considered a medium risk. (3) When the absolute value of the parameter is greater than At that time, it is considered high risk; Step 2: Calculate the risks of runway drift, tail strike, and hard landing respectively. Combining the offset, pitch angle, and vertical landing velocity obtained at various time points, the risk value of each parameter at each time point is obtained through normalization, and denoted as follows: , , The calculation formula is as follows: in, Indicates the first The parameters are in The corresponding risk value, Indicates the first The parameters are in The maximum value at time 1. Indicates the first The parameters are in The minimum value at time; Indicates the first The parameters are in The value at time; Step 3: Construct a comprehensive landing risk model for fixed-wing UAVs A comprehensive landing risk model is constructed by combining the three risks of runway deviation, tail strike, and hard landing. Its expression is as follows: in, To integrate the landing risk model, This represents the risk function that takes the maximum value. This represents the risk function for calculating the mean. Step 3: Modeling the landing sequence of the fixed-wing UAV; Step 1: Construct a local temporal model based on a bidirectional long short-term memory network The processed landing phase data is input into a bidirectional long short-term memory network, and regularization techniques are introduced to prevent overfitting. The calculation process can be expressed as follows: in, This indicates that the feedforward long short-term memory network is in time. The output at that location, This represents the Long Short-Term Memory (LSTM) network function. This indicates regularization techniques. express The probability, This indicates that the inverse long short-term memory network is in time. The output at that location, Indicates vector concatenation. This represents the hidden vector obtained through temporal modeling. Step 2: Construct a global time series model based on Transformer In the Transformer, the output of the bidirectional Long Short-Term Memory network is transformed based on a learned linear transformation. The vectors are transformed into query vector Q, key vector K, and value vector V, which are used to calculate the similarity between each time step in the time series. , , The calculation formula is as follows: in, , , The vectors in the model are respectively , , The corresponding weight matrix, , , For bias terms; The dot product similarity between the query vector Q and the key vector K is calculated and normalized using the softmax function to obtain the attention weights at each time step. These attention weights are then applied to the value vector V using the Attention function to obtain the final self-attention output of the Transformer, calculated as follows: In the formula, This represents the weighted result of self-attention. This indicates the similarity between the query vector and the key vector. The dimension of the key vector; The output based on the self-attention mechanism, after processing through residual structure, layer normalization, and feedforward neural network, yields the final output of the Transformer. The computation process is as follows: in, This represents the intermediate vector after layer normalization. Representation layer normalization; and These represent the outputs of the extended layer and the compressed layer in the feedforward network (FFN), respectively. Indicates the size is The weight matrix, from the input dimension Mapping to Expanded Dimension ; Indicates the size is The weight matrix, from the input dimension Mapping to extended dimensions ; and These represent the bias terms for the FFN expansion and compression layers, respectively. Indicates the encoder layer in time The final output yields the pre-trained prediction model; Step 4: Online risk prediction of fixed-wing UAV landing risk based on streaming learning; Step 1: Calculate the prediction error The actual flight data is input into the pre-trained prediction model for forward propagation, and then the prediction error is calculated and backpropagated. When the prediction error reaches the specified threshold ∆, step 2 is executed; otherwise, step 5 is executed. Step 2: Qualification Update Combining the gradient obtained from backpropagation, a qualification trace is introduced to improve parameter update efficiency and stabilize the model learning process. The qualification trace update process is represented as follows: In the formula, For qualification, The decay factor controls the decay rate of the historical gradient. As a cumulative factor for eligibility traces, it modulates the memory strength in response to sudden perturbations. The gradient at the current time step; Step 3: Dynamically update the learning rate When updating the eligibility trace, the learning rate is updated simultaneously to adapt to unstable data conditions caused by extreme environments. The update formula is as follows: In the formula, for Learning rate at any given moment The basic learning rate, This is the sensitivity coefficient. The magnitude of the state prediction error for a fixed-wing UAV; Step 4: Online update of the prediction model The prediction model is dynamically updated by combining eligibility traces and learning rate, and the parameters in the model are updated using the following formula: in, for Parameter values ​​of the time-based risk prediction model; Step 5: Online Risk Prediction By combining real-time data streams and prediction models, accurate predictions of the landing process status of fixed-wing UAVs in extreme environments are achieved, and online risk prediction is performed based on the comprehensive landing risk model obtained in step two.

Images