An aircraft landing risk prediction method, device and storage medium

Abstract

The application discloses an aircraft landing risk prediction method and device and a storage medium, wherein the method comprises the following steps: acquiring flight parameters, wherein the flight parameters come from a quick access recorder of an aircraft; performing feature extraction processing on the flight parameters, data regularization to a unified second-level time axis, and encoding mapping to obtain flight parameter time sequence representation; the flight parameter time sequence representation is subjected to lagging perception variable interaction calculation to output an interactive enhanced flight state representation; the flight state representation is subjected to high perception time sequence calculation, and radio altitude information is introduced in the calculation to output a long landing risk prediction probability. The device comprises a multi-sampling rate feature coding module, a lagging perception variable interaction module and a high perception time sequence modeling module. The beneficial effect is that the parameters are mapped to a unified second-level time representation, the lagging interaction features of the parameters are explicitly depicted, the radio altitude information is introduced, and a long landing risk probability with better effect is output.

Description

Technical Field

[0001] This invention belongs to the field of flight safety technology and provides a method, device and storage medium for predicting aircraft landing risks. Background Technology

[0002] Aviation safety has always been a core issue for the civil aviation industry. A safety report released by the International Air Transport Association (IATA) shows that the global commercial aviation accident rate will remain relatively low in 2025, but accidents such as tail strikes and runway deviations will still be quite common, indicating that operational safety during takeoff and landing remains a key focus of civil aviation safety management.

[0003] Long landings are a typical unsafe event during the landing phase, usually manifesting as the aircraft's actual touchdown point exceeding the planned touchdown area, reducing the remaining available runway distance and increasing operational risks such as runway overrun and runway deviation. Quick Access Recorders (QARs) continuously record multi-variable, multi-sampling-rate time-series data on aircraft status, control inputs, and environmental conditions, providing a crucial data foundation for long landing risk prediction. In actual landing processes, asynchronous response relationships often exist between different flight parameters. For example, after a change in throttle position, engine speed responds within a few seconds, and changes in airspeed and ground speed exhibit further time skew; there is also a time skew between elevator position changes and pitch and vertical speed responses.

[0004] However, existing flight safety event identification methods based on QAR data mostly rely on general time-series modeling structures, typically focusing on modeling the overall time series characteristics or variable relationships at the same moment, and using resampling or interpolation for time alignment. This approach weakens the local variation information in high-frequency parameters, affecting the characterization of flight state details before touchdown and resulting in poor risk prediction performance. Furthermore, during landing, there may be delayed response relationships between flight parameters such as airspeed, thrust, vertical velocity, and attitude. Existing technologies usually focus on synchronous variable relationships or global sequence feature representations, lacking explicit modeling of the interaction relationships between parameters at different lag times. This makes it difficult to fully characterize the time lag relationships between flight parameters before touchdown, leading to poor risk prediction performance. Moreover, long landing risks are not determined by the flight state at a single moment but are related to the continuous changes in energy state and control actions before touchdown. Existing technologies often focus more on the final classification result or static window-level representation, failing to fully characterize the risk formation and evolution process at different stages before touchdown, resulting in poor risk prediction performance. Furthermore, existing technologies typically only provide risk prediction results, parameter importance, or the importance of time segments, making it difficult to further reveal the lag effects of different parameters at different time offsets. Therefore, they struggle to provide fine-grained evidence for flight quality analysis and pilot training. In summary, existing technologies are insufficient to meet the needs for advance risk prediction and interpretable analysis in complex landing scenarios. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method for predicting aircraft landing risks, which aims to solve the problems in the prior art of QAR data processing distortion, lack of detail characterization, difficulty in fully characterizing the time lag relationship between flight parameters before touchdown, and difficulty in fully characterizing the risk formation and evolution at different stages before touchdown.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting aircraft landing risks, characterized by comprising: The flight parameters are acquired from the aircraft's fast access recorder. The flight parameters are then processed by feature extraction, and the data is regularized to a uniform second-level timeline and encoded and mapped to obtain a time series representation of the flight parameters. The flight parameter time series representation is calculated through hysteresis sensing variable interaction to output an interactively enhanced flight state representation. The flight status representation is performed through high-sensitivity time-series calculations, incorporating radio altitude information to output the probability of long landing risk prediction.

[0007] Furthermore, the acquisition of flight parameters, which are obtained from the aircraft's fast access recorder, involves feature extraction processing of these parameters, data normalization to a uniform second-level timeline, and encoding mapping to obtain a time-series representation of the flight parameters, including: Extract a time length as the observation window; Within the observation window, flight parameters of each category are sampled to form a corresponding set of flight parameters, and each category of flight parameters has a corresponding sampling frequency; The flight parameters are feature-encoded along the original sampling axis of each category of flight parameters; The flight parameters, after feature encoding, are aligned to a unified second-level time axis to serve as a new time series representation of the flight parameters.

[0008] Furthermore, the feature encoding of the flight parameters along the original sampling axis of each category of flight parameters includes: When the sampling frequency is 1Hz, the corresponding second-level features of each flight parameter are obtained directly through linear mapping; When the sampling frequency is greater than 1Hz, the original sampling sequence of each flight parameter within the observation window is expanded in chronological order. Then, along the original high sampling time axis of each flight parameter, one-dimensional convolution combined with activation function is used to extract the local temporal features of the parameter. The extracted features are then uniformly arranged into a standard time step sequence in seconds. The standard time step sequence is then subjected to dual-path aggregation of median pooling and TopK mean pooling. The dual-path aggregation results are then concatenated and linearly mapped to obtain the second-level features corresponding to the flight parameter.

[0009] Furthermore, the time-series representation of the flight parameters undergoes interactive calculation based on hysteresis-sensing variables to output an interactively enhanced flight state representation, including: This study describes the interaction relationships of flight parameter time series representations under different time delays, as well as the degree of dependence of different flight parameter time series representations on various lag relationships at different landing stages.

[0010] Furthermore, the description of the interaction relationships of flight parameter time series representations under different time delays, and the degree of dependence of different flight parameter time series representations on various lag relationships at different landing stages, includes: Construct a multi-offset lag set, pre-configure a lag set containing multiple sets of lag time offsets, and set a corresponding number of lag interaction heads; when the offset value is zero, the corresponding interaction head models the synchronization relationship of different flight parameters in the same second; when the offset value is greater than zero, the corresponding interaction head models the delay relationship between the current time and the flight parameters before the corresponding number of seconds.

[0011] The interaction header generates query, key, and value feature matrices. For each lag interaction header and each second-level time node, the query matrix is ​​generated using the flight code features of the current time, and the key and value matrices are generated using the historical flight code features of the corresponding lag offset time. If there is no valid historical data at the lag offset time, the group of interaction outputs is directly set to zero to avoid interference caused by filling invalid data.

[0012] The lagged variable interaction features are calculated head by head. For each query matrix and key matrix, the scaling dot product similarity is calculated. Then, the value matrix information is fused by normalized weights, and the variable interaction features corresponding to a single lagged interaction head are output.

[0013] Adaptive weighted fusion of all interaction head features is performed. For each second and each type of flight parameter, routing input features are constructed by combining the current encoding state of the parameter, the encoding state of the previous second, and the parameter changes before and after. The adaptive weights of each lag interaction head are calculated. The output features of each interaction head matching the parameter are weighted and concatenated according to the weights. Then, the dimensions are uniformly mapped through a learnable projection layer to obtain the overall fused lag interaction features.

[0014] The residual structure outputs enhanced flight features by performing residual fusion of the fused lag interaction features and the original second-level flight coding features. The fusion result is then normalized by layers and input into the feedforward network. Based on the fusion result, the output of the feedforward network is updated with residuals, and finally, the enhanced flight state coding features that fuse synchronization and delay relationships are output.

[0015] Furthermore, the flight state representation undergoes high-altitude sensing time-series calculation, and radio altitude information is incorporated into the calculation to output the long landing risk prediction probability, including: The flight state representation is vectorized and compressed to obtain a second-level state representation. Arrange the second-level state representations in chronological order to obtain a time state sequence, and then use time series modeling to characterize the time series evolution process within the observation window to obtain the time series hidden state; The encoded features corresponding to the radio altitude parameter RALTC are extracted from the time-series representation of the flight parameters as altitude condition information; The time weights of each time step are calculated based on the altitude condition information and the temporal hidden state to determine the contribution of different time segments to the prediction of long landing risk. The hidden state sequence is weighted and aggregated using the time weights to obtain an observation window-level risk representation. By inputting the observation window-level risk representation into the prediction layer, the predicted probability of long landing risk is obtained.

[0016] The present invention also provides an aircraft landing risk prediction device, comprising: The multi-sampling-rate feature encoding module is used to process flight parameters with different sampling frequencies in the aircraft's fast access recorder and map the flight parameters to a unified second-level time representation to form a flight parameter time series representation. The lag-aware variable interaction module is used to model the dependencies of different flight parameters at multiple time offsets, realize the interaction calculation of lag-aware variables, and thus explicitly characterize the lag interaction features between flight parameters. The altitude-aware temporal modeling module is used to model the temporal evolution of the flight state after the interaction enhancement, and introduces radio altitude information during the time aggregation process to realize altitude-aware temporal calculation, and finally outputs the probability of long landing risk prediction.

[0017] The present invention also provides an aircraft landing risk prediction device, comprising: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to implement the aircraft landing risk prediction method described in any one of the above embodiments.

[0018] The present invention also provides a computer-readable storage medium storing computer instructions for execution by the computer of any one of the above-described aircraft landing risk prediction methods.

[0019] The beneficial effects of this invention are: In aircraft landing risk prediction, by extracting features from flight parameters, normalizing the data to a uniform second-level time axis and completing encoding mapping, and preserving the local dynamic information of high sampling rate parameters, the details of the flight state before landing can be more accurately depicted.

[0020] In aircraft landing risk prediction, the system outputs the synchronization and delay relationships between fused QAR parameters and can adaptively select interaction information under different lag conditions based on the local dynamic state of different parameters, providing an interactively enhanced flight state representation for subsequent high-altitude perception time-series modeling.

[0021] In aircraft landing risk prediction, combining flight state evolution characteristics and altitude stage information helps determine the contribution of different time segments to long landing risk prediction, thereby enhancing the prediction effect.

[0022] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a diagram illustrating the main steps of the method of the present invention; Figure 2 This is a diagram of a sub-step of the method of the present invention; Figure 3 This is a diagram of a sub-step of the method of the present invention; Figure 4 This is a diagram of a sub-step of the method of the present invention; Figure 5 This is a diagram of a sub-step of the method of the present invention; Figure 6 This is a diagram of a sub-step of the method of the present invention; Figure 7 This is a structural block diagram of the aircraft landing risk prediction device of the present invention; Figure 8 A graph showing the changes in key flight parameters for long-landing samples; Figure 9 Distribution of parameter importance in long-landing risk prediction; Figure 10 Temporal importance distribution of long landing risk prediction; Figure 11 Lag correlation analysis of representative QAR parameter pairs. Detailed Implementation

[0024] The present invention will be further described below with reference to specific embodiments. The accompanying drawings are for illustrative purposes only, representing schematic diagrams rather than actual physical objects, and should not be construed as limiting the scope of this patent. To better illustrate the embodiments of the present invention, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0025] like Figure 1-6 As shown, a method for predicting aircraft landing risks includes the following steps: S100. Obtain flight parameters from the aircraft's fast access recorder. Perform feature extraction on the flight parameters, normalize the data to a uniform second-level time axis, and complete the encoding mapping to obtain a time series representation of the flight parameters.

[0026] This step uses the Quick Access Recorder (QAR) as the source of raw flight data and performs calculations on the raw flight data, normalizing the data to a unified second-level timeline. Its technical advantages are: aligning flight parameters with different sampling frequencies to a second-level timeline facilitates subsequent extraction of parameter lag relationships at different time offsets; standardizing data format and dimensions adapts to batch parallel computation with multi-head lag interactions in subsequent calculations, simplifying the algorithm logic; uniformly handling data missing values ​​and dimensional differences reduces noise and ensures the stable operation of the adaptive weighting mechanism; fixing the time granularity accurately depicts the short-term parameter delay correlations during the landing phase, resulting in more accurate modeling; unifying the input and output interfaces of the front and rear modules forms a complete standardized link, facilitating engineering implementation and model training; and retaining local dynamic information of high-sampling-rate parameters provides a more accurate depiction of the flight state details before touchdown, improving the accuracy of risk prediction.

[0027] Specifically, step S100 includes the following sub-steps: S110. Extract a time length as the observation window.

[0028] In this step, for a landing event, an observation window of length T seconds is extracted from the preset landing reference point. The technical advantages of selecting the time length as the observation window are: removing irrelevant data, avoiding interference, limiting data boundaries, unifying sample size, ensuring subsequent calculation efficiency, avoiding noise from irrelevant historical data from distant locations, and limiting the total amount of data for a single sample to reduce storage and computing power requirements.

[0029] S120. Sample the flight parameters of each category within the observation window to form the corresponding flight parameter set, and each category of flight parameters has a corresponding sampling frequency.

[0030] In this step, N QAR parameters are selected as input, where the i-th parameter is represented as: ;in, Let represent the sampling frequency of the i-th parameter, and T represent the observation window length. Different parameters have different sampling frequencies, namely 1Hz, 2Hz, 4Hz, 8Hz, and 16Hz.

[0031] S130. Feature encoding of flight parameters is performed along the original sampling axis of each category of flight parameters.

[0032] In this step, flight parameter features are encoded, and different sampling frequencies require different parameter processing. Specifically: When the sampling frequency is 1Hz, the corresponding second-level features of each flight parameter are obtained directly through linear mapping.

[0033] For a parameter with a sampling frequency of 1 Hz, since only one sample value is contained per second, a second-level feature representation can be obtained directly using a linear mapping: .

[0034] For parameters with a sampling frequency greater than 1Hz, the original sampling sequences of each flight parameter within the observation window are expanded in chronological order. Then, along the original high-sampling time axis of each flight parameter, one-dimensional convolution combined with activation functions is used to extract the local temporal features of the parameter. The extracted features are then uniformly arranged into a standard time step sequence in seconds. The standard time step sequence is then subjected to dual-path aggregation of median pooling and TopK mean pooling. The dual-path aggregation results are then concatenated and linearly mapped to obtain the second-level features corresponding to the flight parameter.

[0035] Specifically, the original sampling sequence within the entire observation window T is first expanded in chronological order as follows: Then, local temporal features are extracted along the original sampling axis using one-dimensional convolution and activation functions, and rearranged in second-level time steps: ; .in, This is represented as the local feature sequence of the i-th parameter. To compress the high sampling rate parameters to a second-level time scale, this step involves... A two-way aggregation process is performed, using median pooling and TopK mean pooling. Median pooling reduces the impact of isolated fluctuations within a second, while TopK mean pooling preserves more significant local dynamic responses. The aggregation results are concatenated and then linearly mapped to obtain the second-level feature representation of this parameter. in, To preserve the proportion for TopK, Indicates feature splicing, This represents the number of local features selected in TopK mean pooling.

[0036] S140. Align the flight parameters after feature encoding to a unified second-level time axis as a new time series representation of flight parameters.

[0037] In this step, through the above processing, all QAR parameters are encoded into a unified second-level feature representation. While preserving the local dynamic information of high sampling rate parameters, it provides a unified time basis for the subsequent interaction module of lag-sensing variables.

[0038] S200. The flight parameter time series representation is calculated through hysteresis sensing variable interaction to output an interactively enhanced flight state representation.

[0039] In this step, lag-aware variable interaction calculations are performed on the time-series representation of flight parameters to model the delay dependencies between different flight parameters on a unified second-level time axis. Since changes in parameters such as thrust, velocity, vertical velocity, and attitude are typically asynchronous during the landing phase, modeling only the variable relationships at the same moment is insufficient to fully reflect the time-dislocation dependencies in the flight state before touchdown. Its technical advantages are: it can model the synchronous coupling relationship of different flight parameters at the same moment, and can also depict the delayed correlation between the current flight state and historical parameters from a few seconds ago, breaking through the limitation of modeling only single-moment parameters, and restoring the time misalignment pattern caused by sensor sampling and aerodynamic response during landing; there are obvious time differences in the responses of physical quantities such as thrust, airspeed, vertical velocity, radio altitude, and attitude angle during the landing phase, and single-moment features cannot reflect the sequential transmission relationship of parameters; by setting multiple hysteresis offsets, it captures the temporal lag effect within a short time scale, which fits the high dynamic flight scenario before touchdown; the output integrates the synchronous and delayed relationships between QAR parameters, and can adaptively select interactive information under different hysteresis conditions according to the local dynamic state of different parameters, providing an interactively enhanced flight state representation for subsequent altitude-sensing temporal modeling.

[0040] Specifically, step S200 includes the following sub-steps: S210. Describe the interaction relationship of flight parameter timing representation under different time delays.

[0041] Furthermore, including: S211. Construct a multi-offset lag set, pre-configure a lag set containing multiple sets of lag time offsets, and set a corresponding number of lag interaction heads; when the offset value is zero, the corresponding interaction head models the synchronization relationship of different flight parameters in the same second; when the offset value is greater than zero, the corresponding interaction head models the delay relationship between the current time and the flight parameters before the corresponding number of seconds.

[0042] Specifically, a lag set is pre-set. ,in, This represents the h-th lag time offset, where H represents the number of lag interaction headers. At that time, this interaction header is used to model the synchronization relationship between different QAR parameters within the same second; when At that time, this interactive header is used to model the current flight state and the previous state. The delay relationship between the flight states in seconds.

[0043] S212. Generate query, key, and value feature matrices for each interaction header. For each lag interaction header and each second-level time node, generate a query matrix using the flight coding features of the current time and generate a key matrix and value matrix using the historical flight coding features of the corresponding lag offset time. If there is no valid historical data at the lag offset time, set the group of interaction outputs directly to zero to avoid interference caused by filling invalid data.

[0044] Specifically, for the h-th delayed interaction header, at the t-th second, a query matrix is ​​first generated using the encoding features of the current time, and then using the delayed time... The encoded features generate the key matrix and value matrix:

[0045]

[0046]

[0047] in, , and These represent the query matrix, key matrix, and value matrix under the h-th lagging interaction header, respectively.

[0048] S213. Calculate the lagged variable interaction features head by head. Calculate the scaling dot product similarity for each query matrix and key matrix. Then, perform weighted fusion of the value matrix information according to the normalized weights and output the variable interaction features corresponding to a single lagged interaction head.

[0049] Specifically, based on the query matrix, key matrix, and value matrix, the variable interaction characteristics corresponding to the h-th lag interaction head are calculated:

[0050] in, This represents the variable interaction characteristic output by the h-th lag interaction head. Indicates the dimension of the key vector. When When the value is less than 1, the corresponding output is set to zero to avoid introducing invalid interactive information caused by padding.

[0051] S220 describes the degree to which different flight parameter time series characterizations depend on various lag relationships at different landing phases.

[0052] Furthermore, including: S221. Adaptively weighted fusion of all interaction head features: For each second and each type of flight parameter, construct routing input features by combining the current encoding state of the parameter, the encoding state of the previous second, and the parameter changes before and after the time interval; calculate the adaptive weight allocation for each lag interaction head; weighted concatenation of the output features of each interaction head matching the parameter according to the weights; and then unified mapping dimension through a learnable projection layer to obtain the overall fused lag interaction features.

[0053] Specifically, the dependence of different QAR parameters on various hysteresis relationships varies at different landing stages. Therefore, this application sets up an adaptive hysteresis weighting mechanism to adjust the contribution of different hysteresis interaction heads based on the local dynamic changes of each parameter. For the i-th parameter at second t, the routing input is constructed based on its current state, the state of the previous second, and the changes in adjacent seconds. And calculate the adaptive weights for different hysteresis interaction heads:

[0054] When t=1, Initialize to a zero vector; weights This represents the weight distribution of the i-th parameter at second t across different lag interaction heads. Subsequently, the output features corresponding to the i-th parameter in each lag interaction head are extracted. And by combining adaptive weights, the fused hysteresis interaction features are obtained:

[0055] in, Let be the learnable projection matrix. Through the above method, the differentiated dependency information extracted by different hysteresis interaction heads is preserved and uniformly mapped back to the original feature dimensions. The fusion result of all parameters and all time steps is denoted as . .

[0056] S222, The residual structure outputs enhanced flight features by performing residual fusion of the fused lag interaction features and the original second-level flight coding features; the fusion result is normalized by layers and then input into the feedforward network, and the output of the feedforward network is updated based on the fusion result, and finally the enhanced flight state coding features of the fused synchronization relationship and delay relationship are output.

[0057] Specifically, a residual update structure is used to fuse the interaction features of lagged variables with the original second-level features:

[0058]

[0059] in, This is the output of the lag-aware variable interaction module. This output integrates the synchronization and delay relationships between QAR parameters on a unified second-level timeline, and can adaptively select interaction information under different lag conditions based on the local dynamic states of different parameters, providing an enhanced flight state representation for subsequent altitude-aware time-series modeling.

[0060] S300. The flight status representation is performed through high-altitude sensing time-series calculation, and radio altitude information is introduced during the time aggregation process to obtain the long landing risk prediction probability.

[0061] In this step, the interaction enhancement features output by the hysteresis-aware variable interaction module are received. This technology is used to model the temporal evolution of flight status during landing and to predict long landing risks by combining radio altitude information. Its advantages are: radio altitude is a crucial parameter characterizing landing conditions and can be used to delineate different landing phases; by combining flight status evolution characteristics and altitude condition information, the contribution of different time segments to long landing risk prediction can be determined; and by aggregating temporal hidden states according to time weights, the accuracy of long landing risk prediction is improved.

[0062] Specifically, S300 includes the following sub-steps: S310. The flight state representation is vectorized and compressed to obtain a second-level state representation.

[0063] Specifically, for the t-th second, take the interaction enhancement features. Then, the features of all QAR parameters at that moment are vectorized and compressed to obtain a second-level state representation: ,in, This is a feature compression mapping unit used to convert multi-parameter interactive features within the current second into a compact state vector.

[0064] S320. Arrange the second-level state representations in chronological order to obtain a time state sequence, and use time series modeling to characterize the time series evolution process within the observation window to obtain the time series hidden state.

[0065] Specifically, the state representations at the second level are arranged in chronological order to obtain a time state sequence. This data is then input into the Mamba time series modeling unit to characterize the temporal evolution process within the observation window. The Mamba time series update process is represented as: ,in, Presentation layer normalization operation, This indicates a hidden time state.

[0066] S330. Extract the encoded features corresponding to the radio altitude parameter RALTC from the flight parameter time series representation as altitude condition information.

[0067] Specifically, since radio altitude can directly reflect the aircraft's altitude changes during landing, the output from the multi-sampling rate feature encoding module... Extract the encoded features corresponding to the RALTC parameters as height conditional information. Let... Let RALTC be the index of the selected QAR parameter, then the height condition at second t is expressed as: .

[0068] S340. Calculate the time weight of each time step based on the altitude condition information and the temporal hidden state to determine the contribution of different time segments to the prediction of long landing risk.

[0069] Specifically, embedding height conditions With time-hidden state Together they are used to calculate the time weights for each time step:

[0070]

[0071] in, This represents the score contributed at time t. This represents the normalized time weights. In this way, the model can combine flight state evolution characteristics and altitude stage information to determine the contribution of different time segments to the prediction of long landing risks.

[0072] S350. The hidden state sequence is weighted and aggregated by time weight to obtain the observation window level risk characterization.

[0073] Specifically, based on the time weights corresponding to each time step, the temporal hidden state sequence is weighted and summed to obtain the observation window-level risk representation: .

[0074] S360. Input the observation window-level risk representation into the prediction layer to obtain the predicted probability of long landing risk.

[0075] Specifically, the observation window-level risk representation g is input into the prediction layer to obtain the predicted probability of long landing risk: ,in, and For learnable parameters, For the Sigmoid function, This represents the predicted probability that the input flight sample belongs to the long landing sample.

[0076] This invention uses a binary classification cross-entropy loss function to optimize the model parameters: ,in, This indicates the actual label.

[0077] refer to Figure 7 As shown, an aircraft landing risk prediction device includes: The multi-sampling rate feature encoding module 10 is used to process flight parameters with different sampling frequencies in the fast access recorder of the aircraft, and map the flight parameters to a unified second-level time representation to form a flight parameter time sequence representation. The lag-aware variable interaction module 20 is used to model the dependency relationship of different flight parameters at multiple time offsets, realize the interaction calculation of lag-aware variables, and thus explicitly characterize the lag interaction features between flight parameters. The altitude-aware temporal modeling module 30 is used to model the temporal evolution of the flight state after the interaction enhancement, and introduces radio altitude information during the time aggregation process to realize altitude-aware temporal calculation, and finally outputs the long landing risk prediction probability.

[0078] An aircraft landing risk prediction device includes: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processors, the instructions being executed by the processors to implement the aircraft landing risk prediction method.

[0079] A computer-readable storage medium storing computer instructions for execution by a computer to implement the aircraft landing risk prediction method.

[0080] Experimental data and results of this invention in specific implementation 1. QAR data This invention evaluates the proposed LAVAN model through two types of verification experiments. The first type is comparative verification, which compares the method of this invention with existing flight safety event prediction methods and time series classification methods to verify the applicability of the method in long landing risk prediction tasks. The second type is ablation verification, which examines the contributions of the multi-sampling rate feature encoding module, the hysteresis-aware variable interaction module, and the altitude-aware time series modeling module to the prediction results.

[0081] The dataset used in this invention contains 8152 QAR samples from the landing phase of a Boeing 737 aircraft, of which 294 are long landing samples and the remaining 7858 are normal samples. Each flight sample records multiple flight parameters during the approach and landing phases. This invention selects 25 key parameters for long landing risk prediction, as shown in Table 1, covering information such as aircraft speed, altitude, attitude, vertical motion, control input, engine status, and environmental factors, with a sampling frequency range of 1 Hz to 16 Hz. The dataset is divided into training, validation, and test sets using stratified sampling in a 6:2:2 ratio. To reduce the impact of class imbalance on model training, only long landing samples in the training set are randomly oversampled, while the validation and test sets maintain their original class distribution. Referring to existing long landing research, this invention uses the distance to the touchdown point as the criterion for long landing labeling: when the horizontal distance between the aircraft at radio altitude 50 ft and the actual touchdown point exceeds 2600 ft, the landing sample is labeled as a long landing sample; otherwise, it is labeled as a normal landing sample. To evaluate the model's predictive capability under different pre-grounding information conditions, for each landing sample, a 32-second truncation of QAR data was used as an observation window, with different prediction endpoints serving as the window's endpoints. Specifically, the prediction endpoints were set to the grounding time, the time at a radio altitude of 10 ft, and the time at a radio altitude of 20 ft, corresponding to scenarios A, B, and C, respectively. During model training, the hysteresis-aware variable interaction module employed hysteresis sets. These are used to model the interaction relationships of variables under different time delays of 1 s, 2 s, and 4 s, respectively.

[0082] Table 1 QAR Parameter Description

[0083] 2. Comparative Experiment To verify the effectiveness of LAVAN in long landing risk prediction, this invention selects seven models—SVM, LSTM, MLSTM-FCN, PatchTST, SDTAN, IMTCN, and TF-QAR—as comparison methods. All models are trained and tested under the same data partitioning conditions, and are evaluated using accuracy, precision, recall, and F1 score.

[0084] Table 2 Performance comparison of different models

[0085] The comparative experimental results are shown in Table 2. LAVAN achieved good prediction results in scenarios A, B, and C, especially in terms of recall and F1 score. Methods such as SVM, LSTM, MLSTM-FCN, and PatchTST can model QAR time series from different perspectives, but they mainly focus on static discrimination, time-dependent modeling, or time segment representation, and their characterization of local dynamic changes and delayed response relationships among multi-sampling-rate flight parameters is still insufficient. In contrast, methods such as SDTAN, IMTCN, and TF-QAR have structural designs for QAR flight safety data and have certain applicability in related flight safety event prediction tasks, but they mainly focus on time feature extraction, local dynamic modeling, and synchronous variable relationship modeling. As the prediction endpoint is advanced from the grounding time to radio altitudes of 10 ft and 20 ft, the terminal landing information available to each model decreases, and the prediction difficulty increases accordingly. Under these conditions, LAVAN still maintains good prediction performance, indicating that the present invention can capture risk-related patterns that appear before grounding, rather than relying solely on discrimination information close to the grounding moment, and has a certain predictive capability.

[0086] 3. Ablation test To further verify the role of each component of LAVAN in predicting long landing risks, this invention constructs corresponding variants for the three components: multi-sampling rate feature encoding, lag-aware variable interaction, and high-aware temporal modeling. First, the local encoding operation in the multi-sampling rate feature encoding module is eliminated, and QAR parameters from different sampling frequencies are directly aggregated onto a second-level time axis. Second, ordinary multi-head attention replaces lag-aware variable interaction, modeling only the variable relationships at the same time point. Third, high-aware temporal modeling is eliminated, and the second-level features enhanced by the interaction are directly aggregated and classified. All of these variants maintain the same data partitioning, training settings, and evaluation metrics as the complete LAVAN.

[0087] Table 3 presents the ablation experiment results. Removing or replacing any of the modules decreased the F1 score of the model in all three scenarios, indicating that each module plays a positive role in predicting long landing risks. The most significant performance decrease occurred when the lag-aware variable interaction module was replaced with a standard multi-head attention mechanism, suggesting that modeling only synchronous variable relationships is insufficient to fully characterize the time delay relationship between flight parameters and long landing risks. Removing the multi-sampling rate feature encoding module also resulted in a performance decrease, indicating that directly aggregating high-frequency parameters weakens the model's ability to express local dynamic changes. Removing the altitude-aware temporal modeling module maintained high precision in some scenarios, but the recall was lower than the complete model, indicating that this module helps capture the temporal evolution characteristics of risk-related states within the observation window and reduces missed detections of long landing samples.

[0088] Table 3 Ablation Experiment Results

[0089] 4. Interpretability Analysis To further illustrate the interpretability of the LAVAN model proposed in this invention, a typical long landing sample was selected for analysis. This sample has a horizontal distance of 3210 ft from the 50 ft position to the actual touchdown point, exceeding the preset long landing threshold of 2600 ft, and is therefore classified as a long landing sample. By visualizing the changes in key parameters, parameter importance, and time weights of this sample, and combining this with lag correlation analysis between parameters, it is demonstrated that this invention can not only provide long landing risk prediction results but also offer explanatory evidence corresponding to changes in flight status.

[0090] like Figure 8 and Figure 9 As shown, the long landing sample exhibited a relatively high speed during the approach phase and later maneuvering corrections before touchdown. Figure 8 (c) shows that the IAS was higher than the normal sample mean for most of the observation window and approached the upper boundary of the normal sample in the later stage, indicating that the aircraft maintained a high approach speed before touchdown. This phenomenon is consistent with... Figure 9 The results for the parameters with high importance in GS and IAS are consistent. Thrust-related parameters also show a clear temporal relationship, among which... Figure 8 In (h), TLA1 decreases in a stepwise manner after approximately 23 seconds. Figure 8 In (g), N11 decreased significantly after approximately 24 s. Simultaneously, attitude, vertical motion, and manipulation-related parameters showed concentrated changes in the later stages. Figure 8 In (e), the pitch is relatively low in the early and middle sections and increases rapidly as it approaches ground. Figure 8 (f) shows a significant fluctuation in PITCH_CMD after approximately 20 seconds; Figure 8 (b) and Figure 8 In (d), ELEVL and ELEVR undergo significant changes in the final stage; Figure 8 In (a), the IVV deviated from the normal range of variation in multiple time periods, and changed rapidly after approximately 22 seconds. This phenomenon is consistent with... Figure 9 The results, which show that parameters such as IVV, PITCH, and PITCH_CMD make significant contributions, indicate that the risks identified by LAVAN do not correspond to a single high-speed state, but are related to multiple flight states such as high approach speed, changes in thrust parameters, changes in vertical motion, and subsequent control-related adjustments.

[0091] Figure 10The time importance results for this sample are presented. As shown in the figure, the model's time weight is relatively low in the early part of the observation window, begins to rise after approximately 22 seconds, and remains at a high level between 26 and 31 seconds, peaking around 28 seconds. This time period is related to... Figure 8 The phases in which parameters such as PITCH, PITCH_CMD, and IVV change significantly are largely consistent. This demonstrates that LAVAN can pinpoint key time intervals relevant to long-landing risk prediction and reflects the phased characteristics where persistently high velocity conditions in the early stages constitute the relevant background, and changes in flight conditions in the later stages significantly contribute to model predictions.

[0092] Furthermore, to explain the rationality of the hysteresis-sensing variable interaction mechanism in this invention from a global perspective, this invention performs hysteresis correlation analysis on representative QAR parameter pairs under different landing results. For example... Figure 11 As shown, for each parameter pair, using the time series of the target parameter as a benchmark, the preceding states of the source parameter are time-aligned with the target parameter according to different hysteresis values, and their correlation within the corresponding landing window is calculated. The selected parameter pairs include TLA1 and N11, N11 and IAS, N11 and GS, ELEVL and PITCH, PITCH and IVV, and IVV and RALTC, covering key relationships in the landing process such as thrust response, velocity changes, attitude adjustment, vertical motion, and altitude changes. The results show that the correlation of different parameter pairs changes with hysteresis, and the correlation distribution between normal landing samples and long landing samples is not entirely consistent. Specifically, the engine state and velocity-related parameter pairs maintain relatively stable correlations under multiple hysteresis values, while the attitude, vertical motion, and altitude-related parameter pairs show more significant changes under different time offsets. These results indicate that there are time-offset-related correlations between different flight parameters before touchdown, providing data-level support for the explicit modeling of cross-parameter hysteresis dependencies in this invention.

[0093] The advantage is that it further reveals the hysteresis relationship of different parameters under different time offsets, providing fine-grained basis for flight quality analysis and pilot training.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting aircraft landing risks, characterized in that, include: The flight parameters are acquired from the aircraft's fast access recorder. The flight parameters are then processed by feature extraction, and the data is regularized to a uniform second-level timeline and encoded and mapped to obtain a time series representation of the flight parameters. The flight parameter time series representation is calculated through hysteresis sensing variable interaction to output an interactively enhanced flight state representation. The flight status representation is performed through high-sensitivity time-series calculations, incorporating radio altitude information to output the probability of long landing risk prediction.

2. The aircraft landing risk prediction method according to claim 1, characterized in that, The acquisition of flight parameters, which are obtained from the aircraft's fast access recorder, involves feature extraction processing of these parameters, data normalization to a uniform second-level timeline, and encoding mapping to obtain a time-series representation of the flight parameters, including: Extract a time length as the observation window; Within the observation window, flight parameters of each category are sampled to form a corresponding set of flight parameters, and each category of flight parameters has a corresponding sampling frequency; The flight parameters are feature-encoded along the original sampling axis of each category of flight parameters; The flight parameters, after feature encoding, are aligned to a unified second-level time axis to serve as a new time series representation of the flight parameters.

3. The aircraft landing risk prediction method according to claim 2, characterized in that, The feature encoding of flight parameters along the original sampling axis of each category of flight parameters includes: When the sampling frequency is 1Hz, the corresponding second-level features of each flight parameter are obtained directly through linear mapping; When the sampling frequency is greater than 1Hz, the original sampling sequence of each flight parameter within the observation window is expanded in chronological order. Then, along the original high sampling time axis of each flight parameter, one-dimensional convolution combined with activation function is used to extract the local temporal features of the parameter. The extracted features are then uniformly arranged into a standard time step sequence in seconds. The standard time step sequence is then subjected to dual-path aggregation of median pooling and TopK mean pooling. The dual-path aggregation results are then concatenated and linearly mapped to obtain the second-level features corresponding to the flight parameter.

4. The aircraft landing risk prediction method according to claim 1, characterized in that, The flight parameter time-series representation is calculated through hysteresis-sensing variable interaction to output an interactively enhanced flight state representation, including: This study describes the interaction relationships of flight parameter time series representations under different time delays, as well as the degree of dependence of different flight parameter time series representations on various lag relationships at different landing stages.

5. The aircraft landing risk prediction method according to claim 4, characterized in that, The description of the interaction relationships of flight parameter time series representations under different time delays, and the degree of dependence of different flight parameter time series representations on various lag relationships at different landing stages, includes: Construct a multi-offset lag set, pre-configure a lag set containing multiple sets of lag time offsets, and set a corresponding number of lag interaction heads; when the offset value is zero, the corresponding interaction head models the synchronization relationship of different flight parameters in the same second; when the offset value is greater than zero, the corresponding interaction head models the delay relationship between the current time and the flight parameters before the corresponding number of seconds. The interaction header generates query, key, and value feature matrices. For each lag interaction header and each second-level time node, the query matrix is ​​generated using the flight code features of the current time, and the key and value matrices are generated using the historical flight code features of the corresponding lag offset time. If there is no valid historical data at the lag offset time, the group of interaction outputs is directly set to zero to avoid interference caused by filling invalid data. The lagged variable interaction features are calculated head by head. For each query matrix and key matrix, the scaling dot product similarity is calculated. Then, the value matrix information is fused by normalized weights and outputs the variable interaction features corresponding to a single lagged interaction head. Adaptive weighted fusion of all interaction head features: For each second and each type of flight parameter, route input features are constructed by combining the current encoding state of the parameter, the encoding state of the previous second, and the parameter changes before and after the time interval; adaptive weights are calculated for each lag interaction head; the output features of each interaction head matching the parameter are weighted and concatenated according to the weights; and then the dimensions are uniformly mapped through a learnable projection layer to obtain the overall fused lag interaction features. The residual structure outputs enhanced flight features by performing residual fusion of the fused lag interaction features and the original second-level flight coding features. The fusion result is then normalized by layers and input into the feedforward network. Based on the fusion result, the output of the feedforward network is updated with residuals, and finally, the enhanced flight state coding features that fuse synchronization and delay relationships are output.

6. The aircraft landing risk prediction method according to claim 2, characterized in that, The flight status representation is calculated using an altitude-sensing time series, and radio altitude information is incorporated into the calculation to output the long landing risk prediction probability, including: The flight state representation is vectorized and compressed to obtain a second-level state representation. Arrange the second-level state representations in chronological order to obtain a time state sequence, and then use time series modeling to characterize the time series evolution process within the observation window to obtain the time series hidden state; The encoded features corresponding to the radio altitude parameter RALTC are extracted from the time-series representation of the flight parameters as altitude condition information; The time weights of each time step are calculated based on the altitude condition information and the temporal hidden state to determine the contribution of different time segments to the prediction of long landing risk. The hidden state sequence is weighted and aggregated using the time weights to obtain an observation window-level risk representation. By inputting the observation window-level risk representation into the prediction layer, the predicted probability of long landing risk is obtained.

7. An aircraft landing risk prediction device, characterized in that, include: The multi-sampling-rate feature encoding module is used to process flight parameters with different sampling frequencies in the aircraft's fast access recorder and map the flight parameters to a unified second-level time representation to form a flight parameter time series representation. The lag-aware variable interaction module is used to model the dependencies of different flight parameters at multiple time offsets, realize the interaction calculation of lag-aware variables, and thus explicitly characterize the lag interaction features between flight parameters. The altitude-aware temporal modeling module is used to model the temporal evolution of the flight state after the interaction enhancement, and introduces radio altitude information during the time aggregation process to realize altitude-aware temporal calculation, and finally outputs the probability of long landing risk prediction.

8. An aircraft landing risk prediction device, characterized in that, include: At least one processor; And a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor to implement the aircraft landing risk prediction method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that are executed by the computer to implement the aircraft landing risk prediction method according to any one of claims 1-6.

Images