Aircraft landing stage short-time abnormal risk early warning and visual simulation adaptive method based on deep learning

Abstract

The invention provides an aircraft landing stage short-time abnormal risk early warning and visual simulation adaptive method based on deep learning, and the method comprises the following steps: synchronizing a multi-source sensor signal, separating the multi-source sensor signal into an impact channel, a wind shear channel and a residual channel, and mapping the multi-source sensor signal into aircraft-runway-wind field graph node features; features are extracted through a space-time diagram convolutional network, and a future trajectory is predicted in combination with a dual-channel attention decoder under physical constraints such as an energy attenuation template; high-risk actions are rapidly inhibited on the basis of real-time risks, the risk probability is calculated through weighted submerged space sampling, graded alarms are output, meanwhile, the alarms drive visual simulation to be adaptively updated, and model online safety iteration is achieved through hard sample mining after landing. According to the method, physical driving modeling, space-time fusion perception and probabilistic risk assessment are adopted, the dependence of a traditional method on an empirical threshold value is overcome, and the early warning precision, the confidence coefficient and the landing safety are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of aircraft landing warning, and in particular to a deep learning-based short-term abnormal risk warning and visual simulation adaptive method during the aircraft landing phase. Background Art

[0002] Existing aircraft landing risk monitoring systems typically follow a three-stage process: multi-source perception - collaborative estimation - and alert display. During the multi-source perception phase, the system first collects real-time aircraft attitude and altitude information along avionics buses such as ARINC / CAN using multiple sensors, including inertial measurement units, radio altimeters, fuselage longitudinal accelerometers, and onboard radar echoes. The collected signals are timestamped using a central reference clock and then resampled to a fixed time interval using linear or cubic spline interpolation algorithms to provide synchronized input for subsequent data fusion. However, due to differences in communication delay and sampling bandwidth between sensor channels, calibration constants or linear interpolation are currently often used to compensate for these differences. This approach fails to completely eliminate hardware and bus-induced delays, resulting in phase shifts in high-frequency impulse signals and deviations from the actual landing loads. Furthermore, fixed-rate resampling strategies are not conducive to high-fidelity fusion of asynchronous, multi-rate signals. During the collaborative estimation phase, extended Kalman filters (EKFs) or complementary filter frameworks are commonly used to fuse multi-source data. During high altitude and cruise phases, the system is primarily based on inertial navigation output. During the approach phase, the filter gain is adaptively adjusted by incorporating residuals from radio altitude and pressure altitude measurements. As the aircraft approaches decision altitude, it switches to landing mode, temporarily increasing the filter bandwidth to simultaneously capture tens-of-millisecond impact peaks and hundreds-of-millisecond wind shear disturbances. However, fixed or segmented bandwidth and gain strategies struggle to strike a balance between the millisecond and hundreds-of-millisecond dynamic scales. Increasing the bandwidth to capture impact peaks degrades low-frequency wind shear tracking capabilities, and vice versa. Furthermore, existing filters lack an online adaptive bandwidth tuning mechanism for multi-band, multi-mutation conditions, making them unable to reliably estimate the drastic frequency changes throughout the landing process. During the warning display phase, the system typically performs a short-time Fourier transform or wavelet decomposition on the fuselage longitudinal acceleration signal to extract the energy peak in the landing gear resonance frequency band. This peak is then compared with the theoretical response curve based on the ground effect and shock strut dynamics model, with the residual value used as the hard landing criterion. Subsequently, characteristic quantities such as peak acceleration, average sink rate, and remaining runway distance are input into a preset threshold table or a lightweight machine learning classifier to achieve multi-level warning output. At the same time, many solutions also use visualization engines to present the predicted trajectory in the form of heat maps, and incrementally update the model parameters through offline playback after the flight. However, methods that rely on a single frequency domain template or empirical threshold are difficult to decouple when the ground effect, tire pressure, and wind field disturbances are coupled with each other, and the false alarm rate increases significantly with environmental uncertainty. Risk assessment also places excessive emphasis on the threshold determination of instantaneous peak or average descent rate. It lacks the ability to predict the energy attenuation of the entire process after landing online, and does not explicitly incorporate model uncertainty into the warning threshold band. In addition, since parameter updates are only performed through offline playback, it is even more difficult to make timely strategy corrections to real-time friction conditions or sudden wind shear within the same landing window.

[0003] Therefore, in view of the above-mentioned phenomena such as insufficient synchronization accuracy, compromise of filtering bandwidth, difficulty in identifying shock-wind shear coupling, and limited real-time adaptability of risk assessment, how to construct a method that can simultaneously maintain high-frequency shock peak fidelity and low-frequency aerodynamic slow-changing tracking during the entire aircraft landing process and provide high-confidence hard landing risk warning in real time under conditions of sudden changes in multiple environmental parameters has become an urgent problem to be solved. Summary of the Invention

[0004] In view of this, the present invention provides a deep learning-based short-term abnormality risk warning and visual simulation adaptive method for the aircraft landing phase, so as to improve the adaptive capabilities of the short-term abnormality risk warning and flight landing visual simulation system during the landing phase, and effectively support rapid response to emergencies.

[0005] To solve the above problems, the present invention provides the following technical solution: a deep learning-based short-term abnormal risk warning and visual simulation adaptive method for aircraft landing phase, comprising the following steps:

[0006] S1. Acquire four raw signals: the inertial measurement unit, radio altimeter, fuselage longitudinal accelerometer, and airborne radar echo. After compensating for the pre-calibrated propagation delay, map each signal to a unified time grid using a cubic spline kernel function. Switch to landing filter parameters at the decision altitude, define a high-frequency impact window based on the gear resonance template to extract the impact component, use a sliding interval operation to obtain the wind shear component, and obtain the residual component through orthogonal projection of the ground effect energy. Write the three-channel signals (impact, wind shear, and residual) along with the timestamp into a ring buffer.

[0007] S2. Receive data from the ring buffer on the flight control main time axis, backtrack a preset number of seconds to form a landing window, map impact, wind shear, and residuals into aircraft node features, wind field node features, and runway node features, and inject real-time runway friction coefficient and wind shear correction weights. Then, extract spatiotemporal correlation features through causal dilation convolution and multi-scale graph convolution, concatenate them into a state vector, and input it into the policy network. The policy network outputs a two-speed maneuver sequence and its probabilities. In the event of vertical kinetic energy exceeding the limit, a negative reward is used to trigger an online update using Adam-SGD with learning rate gate.

[0008] S3. Query the current aircraft model's ground effect function and gear compression curve to form an expected energy attenuation template. Perform orthogonal projection on the convolution-recurrent features to obtain conserved features, which are then mapped to a latent vector. Hadamard weighting is applied to the latent vector based on the policy probability, and the resulting vector is input into a dual-channel attention decoder to obtain the predicted vertical acceleration, sink rate, and wheel load trajectory for the next 2 seconds. The reconstruction error is then calculated by comparing the predicted trajectory with the measured data.

[0009] S4. Generate a soft index gating coefficient based on the real-time runway friction coefficient and wind farm risk score. Apply the gating coefficient to the latent vector to obtain a weighted latent vector. Use the decoder to quickly infer the weighted latent vector and compare it with the predicted trajectory in S3. Map a real-time penalty based on the error and write it back to the policy network.

[0010] S5. Perform Gaussian perturbation sampling within the weighted latent space according to the modified strategy probability. Restore the full time-series trajectory from the sampled latent vectors through a unified decoder. Extract the peak vertical acceleration, window average sink rate, and maximum yaw angle. Calculate the hard landing risk probability and output a confidence interval based on the sampling variance and strategy entropy. Generate a normal, mild, or severe warning message.

[0011] S6. The visual system receives the warning message and the highest-risk trajectory, performs B-spline interpolation at intervals, and completes the body-to-world coordinate conversion. It maps the risk probability to color and transparency, then projects it into the volume shader. It also injects wind volume, runway friction map, and tire contact curve increment instructions into the physics engine. If the sum of the GPU and physics stepping times exceeds a threshold, the volume mesh resolution is automatically reduced.

[0012] S7. After the aircraft speed drops below a preset threshold, the touchdown frame is locked as the synchronization origin. The aligned true value data packets are extracted and written into the incremental learning buffer pool. The vertical impact error is calculated, hard samples are selected, mixed with an equal number of random samples, and replayed into the policy network. Rapid adaptation is performed according to the weighted gradient of the hard samples. At the same time, the terminal encoding layer of the variational network is unfrozen and the dual-channel attention is fine-tuned with limited iterations. The energy conservation difference and damping ratio error are monitored in real time. If the threshold is exceeded, the weights are immediately rolled back, and the new weights are hot-swapped online after batch replay verification is completed in the isolated sandbox.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] This invention addresses the differences in sampling rates and communication delays among multiple sensor sources by compensating for known delays and synchronously resampling with a cubic spline kernel function. It then dynamically switches to a landing-specific complementary filter at the decision altitude. Furthermore, it extracts high-frequency impact windows based on a gear resonance template, separates low-frequency wind shear features in parallel using a sliding window, and uses orthogonal projection of ground effect energy to remove predictable components. This results in physically driven separation and real-time fusion of the impact, wind shear, and residual signals.

[0015] To overcome the challenges of heterogeneous and sparsely correlated flight-runway-wind field features, the three-channel data are mapped to aircraft, runway, and wind field nodes respectively. Causal dilation convolution is used along the time dimension at the aircraft node to extract dynamic trends. At the same time, multi-scale graph convolution is used to fuse spatial correlation features with the dynamic adjacency matrix generated based on the real-time runway friction coefficient and wind field risk score. The structure completes spatiotemporal multimodal relationship learning within a single tensor, effectively improving the consistency of scene perception.

[0016] Considering the lack of physical constraints in traditional prediction networks, we construct an expected template for ground effect and gear compression energy attenuation. We then map the features extracted by the convolutional-recurrent network into an energy-conserving subspace via orthogonal projection. We then inject the policy network's probability weights into the resulting latent vector, locking the latent space within a physical upper bound. A dual-channel attention decoder applies separate focusing mechanisms for the impact peak and aerodynamic ramp, ensuring that the predicted trajectory faithfully reproduces the dynamics of the three stages: impact, cushioning, and rolling.

[0017] To address the real-time limitations of online risk suppression, a soft exponential gating coefficient is calculated based on the current friction and wind field scores. This is applied to the latent vector and then fed into a fast decoder with frozen structural parameters. This decoder completes forward inference in 10ms, aligns the reconstructed error point-by-point with the original prediction, and maps it into an immediate negative reward write-back policy network, enabling rapid suppression of high-risk actions without requiring additional gradient calculations.

[0018] Weighted Gaussian perturbation sampling is performed within the weighted latent space according to the modified policy probability. A unified decoder then incorporates landing configuration and tire pressure constraints during the decoding process to restore the full time-series trajectory. Peak vertical acceleration, average sink rate, and maximum yaw angle are extracted for each sample. The sampling variance and policy entropy are combined to calculate risk probability and confidence intervals, avoiding reliance on a single threshold and supporting graded alerting.

[0019] The visual simulation process has been adaptively expanded. After receiving warning signals and the most dangerous trajectory, spatial alignment is achieved through B-spline interpolation and aircraft-to-world coordinate transformation, mapping risk probability and uncertainty to the color and transparency of the volume shader. The physics engine simultaneously loads wind field volumes, runway friction maps, and tire contact curve increments, and dynamically adjusts the mesh resolution based on rendering and physics stepping times to ensure highly reliable real-time rendering.

[0020] To address model aging and sample distribution drift, after touchdown frame lock, real-world observation data is aligned. Hard samples are selected based on vertical impact error and mixed with random samples before being replayed back into the policy network and variational network terminal encoding layer. Fast fine-tuning is achieved using hard sample weighted gradients and restricted Adam-SGD, while energy conservation and damping ratio are monitored in real time. Over-threshold rollbacks are initiated, and batch replay verification and hot swapping are completed in a sandbox, ensuring safe and efficient adaptive iteration of the model after landing. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flowchart of the deep learning-based short-term abnormal risk warning and visual simulation adaptive method for aircraft landing phase provided by the present invention;

[0022] Figure 2A logic diagram for constructing landing window feature coding for the deep learning-based short-term abnormal risk warning and visual simulation adaptive method for aircraft landing phase provided by the present invention; DETAILED DESCRIPTION

[0023] The present invention is further described below with reference to specific embodiments.

[0024] like Figure 1 and Figure 2 As shown, the present invention provides a method for short-term abnormal risk warning and visual simulation adaptation during the landing phase of an aircraft based on deep learning, comprising the following steps:

[0025] S1. Acquire four raw signals: the inertial measurement unit, radio altimeter, fuselage longitudinal accelerometer, and airborne radar echo. After compensating for the pre-calibrated propagation delay, map each signal to a unified time grid using a cubic spline kernel function. Switch to landing filter parameters at the decision altitude, define a high-frequency impact window based on the gear resonance template to extract the impact component, use a sliding interval operation to obtain the wind shear component, and obtain the residual component through orthogonal projection of the ground effect energy. Write the three-channel signals (impact, wind shear, and residual) along with the timestamp into a ring buffer.

[0026] Among them, S1 includes the following steps:

[0027] S11. Timestamp synchronization and unified resampling. To eliminate the differences in sampling rate and inherent communication delay among the inertial measurement unit (IMU), radio altimeter, fuselage longitudinal accelerometer, and airborne radar echo, a unified time grid {t k |k=1,…,K};

[0028] Assume that the original time stamp sequences of the IMU, radio altimeter, fuselage longitudinal accelerometer and airborne radar echo are T I ,T H ,T B ,T R , the corresponding observation vector is x I,p ,x H,p, ,x B,p ,x R,p , where the subscript p represents the respective original sampling points;

[0029] The fixed propagation delay compensation pre-calibrated for each sensor is denoted as τ I , τ H ,τ B ,τ R ;

[0030] Perform cubic spline kernel interpolation mapping on the uniform time grid to obtain the aligned observation vector y(t k ), the mapping relationship satisfies:

[0031]

[0032] Among them, s∈{1,2,3,4}, corresponding to IMU, radio altimeter, fuselage longitudinal accelerometer, and airborne radar echo in sequence; N s is the number of original samples in the s-th path; κ s (·) is the cubic spline kernel function designed for the dynamic bandwidth of the road; e s is a 4×1 basis vector used to write four signals into different channels independently of each other; s,p is the observation value of the s-th sensor at the p-th original sampling point; τ s Fixed propagation delay compensation for each channel to ensure that the alignment result y(t k ) does not have phase drift;

[0033] S12. Decision height detection and dedicated filter channel switching, in real-time stream y(t k ) to extract the radio height H(k)=y H (t k ) and vertical descent rate V z (k) = (H(k) - H(k-1)) / Δt;

[0034] Once it is detected that H(k)≤H D (where H D The decision height threshold for joint certification of aircraft type and runway is set as the threshold). The high-altitude filter used in the cruise phase is immediately turned off, and the landing-specific complementary filter channel is activated instead. The ground effect reference model g(k) is loaded from the three-dimensional index table of aircraft type, mass and flap angle.

[0035] All subsequent calculations will be based on the output of this dedicated channel to ensure that the high-altitude filter parameters fully match the dynamic characteristics of the near-ground cushion.

[0036] S13. Gear resonance matching and shock high-frequency window calibration, for the aligned observation vector y of the fuselage longitudinal acceleration B (t k ) online calculation of its short-time Fourier transform power spectrum P(k,f), and retrieval of the resonance template Ψ(f) under the current tire pressure and load conditions from the landing gear mechanical feature library;

[0037] Solve the cross-correlation C(t k ), and then C(t k ) back-calculated to the time domain high-frequency impact window w c , in w c Only keep y B (t k ) high frequency component, and obtain the calibrated shock pulse ic (k), other frequency bands are in a controlled smoothing state in the subsequent filtering architecture to ensure that the transient impact energy generated by gear compression will not be mistaken for noise and attenuated;

[0038] In S11, the fuselage longitudinal acceleration signal has been aligned to the uniform time grid {t k}, recorded as y B (t k ), S13 first performs a short-time Fourier transform (STFT) on the signal of each sliding window length L1 (e.g., 100ms-200ms):

[0039] P(k,f)=STFT{y B (t k )},P(k,f)=|A(k,f)| 2 ;

[0040] Among them, P(k,f) is the short-time Fourier transform power spectrum;

[0041] According to the current tire pressure and load conditions, the frequency domain template Ψ(f) is taken from the gear mechanical feature library and the frequency domain template Ψ(f) is taken from the gear mechanical feature library. k Calculate the cross-correlation between the power spectrum and the template:

[0042]

[0043] The upper and lower limits of the integral are (f min ,f max ), peak point Corresponding to the central moment of a landing impact;

[0044] The high-frequency impact window is defined using the empirical window width Δt and the template bandwidth Δf as parameters:

[0045] w c =[t d -Δt,t d +Δt];

[0046] And perform bandpass filtering of the center frequency and bandwidth Δf on the original acceleration within this time domain area:

[0047] i c (t) = BandPass{y B (t),f ± Δf},t∈w c ;

[0048] The above time domain impulse signal i c (t) According to the unified time grid {t k}Interpolate or resample to obtain a discrete sequence:

[0049] ic (k)=i c (tk);

[0050] In this way, c (k) is the impact pulse data after S13 template matching positioning, high-frequency bandpass filtering, and remapping to a unified grid;

[0051] S14. Low-frequency wind shear feature separation, for the area outside the window Ω={t k}\W c , parallel calculation of IMU lateral acceleration within the sliding interval of length M The velocity difference ΔV(t k )=y R (t k )-y R (t k-1 ), the index is obtained by offsetting the antenna-INS installation error based on the correction coefficient β

[0052] If |Λ(k)|≥σ Λ ,σ Λ As the threshold, the wind shear event is marked within Ω and the wind shear feature sequence c is generated. w (k);

[0053] The operation is naturally mutually exclusive with the impact window in the time dimension, preventing confusion between low-frequency wind shear changes and high-frequency vertical impacts;

[0054] S15. Based on the energy conservation principle, the ground effect residual denoising is performed. The ground effect reference model g(k) is injected into the observation matrix of the Kalman-complementary fusion device, and the longitudinal acceleration a(k)=y is calculated within the current analysis window length M. B (t k ) Perform energy orthogonal projection and calculate the residual:

[0055]

[0056] Where the numerator is the inner product of a(k) and g(k), the denominator is the energy of g(k) within the window, and the entire fraction is the optimal scaling factor in the least squares sense. The ground effect reference model g(k) is a state-of-the-art, pre-established mathematical or empirical model used to describe the specific effects of aerodynamic changes on an aircraft flying close to the ground.

[0057] In step S15, the existing model is introduced to help the Kalman-complementary fusion processor more accurately identify and separate the signal components generated by the ground effect, thereby providing a benchmark or correction basis for subsequent residual denoising;

[0058] In this way, the filter gain is maximized while strictly satisfying the energy conservation constraint, and only the high-frequency residual R(k) that cannot be explained by the ground effect model is retained, effectively suppressing sensor noise.

[0059] S16. The shock pulse i calibrated by S13 c (k) Wind shear characteristic sequence c separated by S14 w (k) is concatenated with the residual R(k) obtained by S15 in ascending order of time index to form a three-channel fusion signal F(k) = [i c (k),c w (k), R(k)], F(k) is written to the shared address BUF1 through the double-ended ring buffer and appended with the complete timestamp table t k ;

[0060] S2. Receive data from the ring buffer on the flight control main time axis, backtrack a preset number of seconds to form a landing window, map impact, wind shear, and residuals into aircraft node features, wind field node features, and runway node features, and inject real-time runway friction coefficient and wind shear correction weights. Then, extract spatiotemporal correlation features through causal dilation convolution and multi-scale graph convolution, concatenate them into a state vector, and input it into the policy network. The policy network outputs a two-speed maneuver sequence and its probabilities. In the event of vertical kinetic energy exceeding the limit, a negative reward is used to trigger an online update using Adam-SGD with learning rate gate.

[0061] Among them, S2 includes the following steps:

[0062] S21. Receive and align the three-channel fusion sequence, on the flight control main axis The three-channel fusion signal F(t k )=[i c (t k ),c w (t k ),R(t k )], K in the flight control main time axis represents the total number of sampling points on the main time axis, and F is the three-channel fusion signal function in S16;

[0063] F(t k ) is mapped to the main time axis through linear interpolation to obtain the alignment result

[0064] The current interpretation time t now Backtrack L seconds for the right endpoint to build a continuous landing window Where N = L / Δt, k0 is the current judgment time t now The index in the timestamp sequence;

[0065] S22. Multi-mode dynamics-environment initial feature map, in the window Internal pair Channel analysis, using the aircraft airspeed meter to read the mass m a ;

[0066] The vertical velocity is obtained by first-order difference Wheel load obtained from main wheel strain gauge Obtaining the crosswind component from the inertial navigation lateral velocity and the meteorological blade meter data

[0067] The three observations are mapped into time series features of aircraft nodes, runway nodes, and wind farm nodes. To maintain dimensional consistency within the graph network, each feature is scaled by the aircraft model normalization scaling factor vector e = [s1,…,s5], and finally a node tensor is constructed:

[0068]

[0069] The symbols are explained as follows:

[0070] X u,j,l : the eigenvalue of the u-th node, time j, and l-th dimension in the graph network;

[0071] u∈{1,2,3}: node type index—1=aircraft, 2=runway, 3=wind farm;

[0072] j=1,…,N: time index, the total number of frames in the lookback window is N;

[0073] l: Feature dimension number, corresponding to the five observations one by one, and other positions are filled with 0;

[0074] e = [s1, s2, s3, s4, s5]: normalized scaling coefficient vector for dimension matching;

[0075] Vertical speed at the jth frame of the flight control clock (first-order difference of the airspeed indicator of the aircraft model);

[0076] The main wheel load at frame j (measured by strain gauge);

[0077] μ w : real-time runway slipperiness coefficient (friction coefficient);

[0078] P t : Real-time tire inflation pressure;

[0079] Crosswind component of the jth frame (fusion of inertial navigation side speed and meteorological blademeter);

[0080] The jth frame time after the flight control main clock is aligned;

[0081] The above structure solves the problem of heterogeneous and sparse flight-environment-runway features, enabling graph convolution to complete multimodal relationship learning within a single tensor;

[0082] S23. Dynamic weight coupling of sparse environmental attributes, based on the μ uploaded in real time by the apron weather radar and runway friction measurement device. w With P t , the second and third columns of the runway node feature matrix are written immediately, and then passed through the weight function W p (k)=α1μ w +α2 / P t Generate an impact weight correction term, which will be used together with the vertex metric to adjust the edge weight in the dynamic adjacency matrix A(k) to achieve dynamic compensation for the impact amplification on wet runways;

[0083] S24. Temporal convolution - multi-scale graph convolution joint state encoding, input the aircraft node along the time dimension into the causal expansion convolution network to obtain the time feature h T (k);

[0084] The runway nodes and wind farm nodes are fed into a three-layer multi-scale graph convolution together with the dynamic adjacency matrix A(k) constructed by S23, and the spatial correlation features h are extracted using a two-order approximation. G (k);

[0085] In the fusion layer, h T (k) and h G (k) performs channel-level splicing and outputs the complete game state vector s(k) = h T (k)‖h G (k)], providing the policy network with a high-dimensional input that is consistent in both time and space;

[0086] The dynamic adjacency matrix A(k) in S23 is composed of the basic topology and the dynamic weights of the environmental attributes. Assuming that the three types of nodes are 1 = aircraft, 2 = runway, and 3 = wind farm, the basic adjacency matrix is ​​as follows:

[0087]

[0088] Among them, the diagonal ring is 1, the aircraft Runway, runway The wind farm is connected, and the rest are 0;

[0089] The dynamic weight increment is as follows:

[0090]

[0091] Δa 12 (k) = Δa 21 (k) = g r (μ w(k)),Δa 23 (k) = Δa 32 (k) = g w (φ w (k)),

[0092] The symbols are explained as follows:

[0093] μ w (k): runway friction coefficient at time k;

[0094] φ w (k): wind farm risk score at time k;

[0095] g r ,g w : A weight mapping function that can be learned or empirically fitted;

[0096] The dynamic adjacency matrix is ​​as follows:

[0097] A(k)=A 0 +ΔA(k);

[0098] That is, for aircraft Runway, runway The edge weights of the wind farm are used for real-time amplification / attenuation of environmental perception, while other locations remain disconnected or self-loop constants;

[0099] S25. Hard landing event detection and online reward update, calculating vertical kinetic energy based on the aircraft inertia parameters at each time step Assume the maximum allowable kinetic energy threshold E lim , if E k (k)≥E lim , it is considered a hard landing and rewards are generated immediately:

[0100]

[0101] Among them, λ>0 is the scale coefficient, η>1 is used to perform high-order amplification of the excess kinetic energy, and ρ is the slipperiness gain;

[0102] This design incorporates the penalty of vertical kinetic energy and runway friction margin simultaneously, reflecting the high-risk characteristics of the wet runway-high kinetic energy combination;

[0103] S26. Policy network adaptive evolution and maneuver sequence generation. The policy network receives the previous moment's parameters and the new reward r(k). Within the same computational graph, it completes one iteration using an Adam-SGD hybrid optimization with learning rate gating to avoid compromising online real-time performance. The updated network generates two maneuver action sequences under the current state s(k): a short-term sequence π1 with a length of T1 = 3 seconds, and a medium-term sequence π2 with a length of T2 = 5 seconds. It also outputs the corresponding probabilities p1(k) and p2(k), thereby balancing the control requirements for different remaining glide distances.

[0104] S27. The risk assessment head outputs a priori risk set, and each candidate sequence π i Through rapid simulation of the aircraft-runway-wind field hybrid dynamics model, the vertical kinetic energy peak is calculated separately. and residual friction margin

[0105] Constructing risk scores based on dual indicators Among them, w1 and w2 are experience weights, which eventually form the risk score set {ρ1, ρ2} and the policy probability {p1(k), p2(k)}, which are output in parallel and provide the power-environment consistency constraint for the prior conditions of S3.

[0106] S3. Query the current aircraft model's ground effect function and gear compression curve to form an expected energy attenuation template. Perform orthogonal projection on the convolution-recurrent features to obtain conserved features, which are then mapped to a latent vector. Hadamard weighting is applied to the latent vector based on the policy probability, and the resulting vector is input into a dual-channel attention decoder to obtain the predicted vertical acceleration, sink rate, and wheel load trajectory for the next 2 seconds. The reconstruction error is then calculated by comparing the predicted trajectory with the measured data.

[0107] Among them, S3 includes the following steps:

[0108] S31. Align and crop the prediction window. The flight control computer advances with a 1000 Hz master clock at the end of landing. When the real time corresponding to the current frame number is t now When it is determined to be in the radar altitude range of 2-3m before the aircraft touches down, the system synchronously receives the three-channel fusion signal F(t k )=[i c (t k ),c w (t k ),R(t k )] and the strategy probability p(k) output by S2, in order to ensure the consistency with the execution delay of elevator-autothrottle, first calculate F(t k ) and the main time axis Do millisecond interpolation alignment, and then use T P = 3s window Intercept data backward, Tp Specifically refers to the length of the prediction / lookback window. The three-second window corresponds to the entire process of the pilot completing the leveling, descent, and touchdown, including the three typical dynamic stages of ground effect lift, initial tire contact, and full gear compression. This constitutes the only input for latent space reasoning.

[0109] S32. Transient-gradient dual feature extraction, in the window Internal construction tensor U k , whose first row i c (t k ) reflects the main wheel rigid impact peak, the second row c w (t k ) describes the low-frequency lateral shear driven by the crosswind shear, and the third row R(t k ) records sensor residual noise. For the spectral differences of impact (short pulses on the order of milliseconds) and wind shear (slowly varying on the order of hundreds of milliseconds), a one-dimensional convolution is first performed with a learnable convolution kernel W1 of length 7 to enhance the high-frequency components and highlight the narrow peak caused by the gear pressure drop. The convolution result is then progressively input into a gated recurrent unit, which uses a recurrent gate to accumulate the crosswind slope over a time constant of 250-500ms.

[0110] Among them, the convolution-recurrence coupling satisfies:

[0111] H k =GRU(σ(W1*U k +b1),H k-1 );

[0112] Where σ = ELU, the physical property of symmetrical stretching corresponding to the impact amplitude, and b1 is the bias;

[0113] After this treatment, H k Expressing the dual characteristics of transient stiffness and gradual aerodynamics, laying a baseline for subsequent energy template comparison;

[0114] S33. Build the expected energy attenuation template. This application aims to further quantify how the energy after impact is distributed among the ground effect air cushion, shock absorber strut, and tire deformation. Query the aircraft model-weight-flap angle index table to obtain the current angle of attack A. α , lower washing angle A ω The dimensionless ground effect function g(k,A α ,A ω );

[0115] Then retrieve the gear compression curve from the tire pressure-oil pressure-temperature triplet and calculate the instantaneous stiffness-damping coefficient C st ;

[0116] The two parts are combined into the expected energy decay template:

[0117] T(k)=C st ·g(k,Aα ,A ω );

[0118] Dimensionless ground effect function g(k,A α ,A ω ) is a predefined mathematical function or a lookup table that can be used to quantify the aircraft's angle of attack A. α and lower washing angle A ω Defines the degree of influence of ground effect on its aerodynamic characteristics under a specific flight attitude and altitude indirectly represented by k or other altitude information indexed by k;

[0119] Instantaneous stiffness-damping coefficient C st A parameter that measures the ability of a landing gear system consisting of shock struts and tires to dissipate vibration energy. The larger the damping coefficient, the faster the vibration decays after an impact, and the greater the energy dissipation.

[0120] Based on this template, the maximum impact kinetic energy E is given in physical terms. imp To stable rolling action E roll The theoretical dissipation trajectory of , which is directly used as a reference for subsequent regularization;

[0121] S34. Physical constraint suppression and latent space consistency, if the convolution-recurrent feature Z(k) = H k At a certain moment, the energy exceeds the template, i.e. The system will rebound without reason or encounter unrealistic secondary wheel contact during the prediction phase. To this end, the excess energy is eliminated through orthogonal projection to obtain the conservative characteristic Z sup (k);

[0122] This ensures that even if the network tends to amplify the impact, it cannot exceed the damping upper bound given by the configuration file, thus strictly locking the latent space within the realistically achievable flight-ground coupling range;

[0123] S35. Inject the strategy probability conditional latent vector into the strategy probability p(k) to reflect the pilot-autopilot mixed intention. If p1(k) (three-second maneuver) is high, it means that the system predicts that the pilot prefers a quick correction; if p2(k) (five-second maneuver) is high, it means that the pilot accepts a longer buffer.

[0124] In order to allow the prediction network to adaptively amplify or weaken attention to different intentions, this step linearly maps p(k) to generate weight α(k) and applies it to Z sup (k) Perform Hadamard weighting to generate the latent vector z(k) = α(k)⊙Z sup (k) When the probability of high-risk actions increases, the variance of the relevant latent dimension increases, guiding the downstream decoder to focus on potential hard landing paths;

[0125] S36. Dual-channel attention decoding prediction trajectory, the decoder first constructs the impact channel attention Att for z(k) a , whose attention is drawn from the gear pressure drop reference signal, specifically capturing the strong nonlinear peak of <20ms;

[0126] Secondly, construct the aerodynamic slowly varying channel attention Att b , its attention is based on the ground effect air cushion attenuation time constant, focusing on the 200-500ms slip trend, the two-channel output is gated and fused to generate the vertical acceleration, sinking rate, and original predicted trajectory for the next 2 seconds This combination ensures that the gear peak can be rebuilt and the transition to the stable rolling phase can be smooth, fitting the three-stage physical process of impact-buffer-rolling.

[0127] S37. Reconstruct the error and the posterior output, and decode the original prediction trajectory Measured with the window Perform the two-norm difference to obtain the reconstruction error ε(k). The instantaneous high value of the error indicates that the model is not aware of the current working conditions, corresponding to the anomaly that is not explained by the physical regularization and intention conditions. In parallel, the posterior mean μ of the latent variable is calculated. z (k) and variance The potential risks can be re-weighted by Bayesian method in the future. The data is written into the shared buffer BUF2, providing a comprehensive input for the gate-fine-tuning closed loop, including both data-driven anomalies and physically interpreted residuals, to achieve precise early warning of extreme hard landing scenarios.

[0128] S4. Generate a soft index gating coefficient based on the real-time runway friction coefficient and wind farm risk score. Apply the gating coefficient to the latent vector to obtain a weighted latent vector. Use the decoder to quickly infer the weighted latent vector and compare it with the predicted trajectory of S3. Map a real-time penalty based on the error size and write it back to the policy network, achieving 10ms-level policy probability suppression.

[0129] Among them, S4 includes the following steps:

[0130] S41. Time alignment of strategy probability and latent variables and buffer writing. In the 500ms period before the aircraft's main wheels touch down, the flight control master clock enters a high-risk, high-decision-density phase. Any strategy delay during this period will directly affect the main wheel pressure drop peak and even the wing torsional margin. Therefore, the strategy probability sequence [p(k) = p1(k), p2(k)] generated by S2 and the posterior latent variable z(k) output by S3 are read simultaneously.

[0131] The two sets of data come from the graph convolution-strategy GPU stream and the variational RNN CPU stream, respectively, and have sub-millisecond arrival jitter. To maintain synchronization with the 16ms update cycle of the elevator actuator, p(k) is first aligned to the latent variable timestamp through linear interpolation. Then follow the index Write to the shared buffer BUF3, where p in the index represents the latent variable and n represents the nth one;

[0132] S42. Hazard preference gate coefficient generation: real-time risk score ρ given by wind farm LiDAR and black surface friction vehicle at 500m from the runway end n (k) and the strategy probability p(k) jointly determine the degree of dynamic stress. In order to make it clear that hard landing has higher priority than normal rollout, the gating coefficient is constructed based on the soft exponential mapping:

[0133]

[0134] Among them, κ is calibrated online based on the current tire-track friction coefficient. w As the aircraft descends, κ automatically increases, further amplifying the weight of high-risk actions in the gating. This ensures that even if the pilot tends to push the stick for a slow descent on a slippery runway, the overall focus will be on the potential consequences of a hard landing. n represents the nth gating channel signal.

[0135] S43. Gated modulation forms a weighted latent space, multiplying g(k) element-wise by the latent variable z(k) to generate a weighted latent vector This operation is equivalent to increasing the impact-ground effect interaction dimension in the latent space and decreasing the normal rolling dimension. When the sudden rise of the crosswind causes the main wheel load to increase sharply, The component related to vertical stress is amplified by the gating coefficient, providing clearer hazard features for subsequent decoders;

[0136] S44. Weighted latent space decoding and reconstruction error evaluation, Input the fast decoder with the same structure as step S3 but frozen parameters to obtain the weighted prediction trajectory The trajectory focuses on restoring the dual-stage curve of the impact peak of 0-50ms and the damping attenuation of the shock absorber strut from 50-300ms, and compares it with the original prediction of S3. Do point-by-point alignment;

[0137] If the gating is effective, the difference between the two trajectories increases near the impact peak and gradually converges in the slow-changing section, thus laying a physical basis for the error-driven policy callback.

[0138] In S44, two decoders with the same structure but different purposes are defined:

[0139] The main decoder Dec is used for online updates during training. The fast decoder Dec0 has the same structure as Dec, but the parameters θ are frozen and only forward reasoning is performed. The following is the original prediction trajectory:

[0140]

[0141] The Hadamard product of the gating coefficient α(k) and the latent vector z(k) is:

[0142] z w (k)=α(k)⊙z(k);

[0143] The weighted prediction is:

[0144]

[0145] Among them, θ is exactly the same as the main decoder and remains frozen, and does not participate in the gradient calculation;

[0146] Finally, the reconstruction error is calculated point by point:

[0147]

[0148] The design ensures that forward reasoning and error calculation are completed within a 10ms latency, providing feedback for the policy network to suppress the error immediately.

[0149] S45. Error-penalty mapping and writing back to the policy network. To immediately suppress uncertain and dangerous actions, the reconstructed error vector is mapped into a real-time penalty:

[0150]

[0151] Among them, λ is adaptively adjusted based on the aircraft model's decision altitude minus the remaining runway length. High-error and high-gating actions will receive the largest negative reward, which is directly written back to the corresponding branch of the policy network, reducing the probability of taking this action in the subsequent 10ms.

[0152] S46. Quickly fine-tune and update the policy probability in the same graph. The policy network uses Δr in the same computation graph. n (k) Perform a restricted-step Adam-SGD update, completely avoiding the repeated forward-backward overhead;

[0153] Since fine-tuning only affects the output layer weights, multi-core ARM operations can be completed within 2ms;

[0154] Generate revised strategy probability immediately after update Reflects the latest decision-making trends of the three-party collaboration between pilot, autopilot and risk gate;

[0155] S47. The system will correct the probability Weighted latent vector and gated risk score Frame the message into a policy latent space-risk ternary message and write it into the shared buffer BUF4;

[0156] Once the message is written, it triggers the start of encrypted sampling of high-risk actions, providing a real-time updated prior distribution for subsequent probability-confidence assessment. This implements a closed-loop control system of hazard identification, strategy suppression, and probability update, significantly reducing the frequency of excessive maneuvers under high-incidence hard landing conditions.

[0157] S5. Perform Gaussian perturbation sampling within the weighted latent space according to the modified strategy probability. Restore the full time-series trajectory from the sampled latent vectors through a unified decoder. Extract the peak vertical acceleration, window average sink rate, and maximum yaw angle. Calculate the hard landing risk probability and output a confidence interval based on the sampling variance and strategy entropy. Generate a normal, mild, or severe warning message.

[0158] Among them, S5 includes the following steps:

[0159] S51. After the flight control main clock detects that the main wheel pressure drop curve has crossed the 80% touchdown threshold, it immediately reads the corrected strategy probability output by S4. With weighted latent variables

[0160] According to the aircraft type - runway length table, determine the remaining maneuvering time of three seconds and five seconds, respectively, and establish windows and Among them, the window and The left endpoint is aligned with the current master clock frame t0, and the right endpoints are located at t0+3s and t0+5s respectively;

[0161] The operation ensures that all subsequent risk estimates cover the entire process from touchdown to front wheel lift to initial roll, which is the minimum complete time domain for ground accident prediction;

[0162] S52. Probabilistic weighted latent space drawing, further to fully cover the heavy tail action, according to the modified strategy probability to perform weighted drawing: each time from the discrete distribution Select action number n∈{3s,5s} from the corresponding latent variable subset Gaussian perturbation is applied to Generate sample z (s) , the weight-perturbation joint sampling satisfies:

[0163]

[0164] Where M is the number of samples, σ z Automatically set based on the S3 posterior variance to ensure that the proportion of high-risk actions in the sample pool is not less than 70%, avoiding over-smoothing. is the diagonal covariance matrix, I is the identity matrix, z (s) =v represents the sth latent sample value v;

[0165] S53, z (s)Input the airworthiness-approved unified decoder D(·) to restore the full time sequence trajectory. The formula is as follows:

[0166]

[0167] Among them, Y (S) Meaning of (t): Y here (S) (t) is a more complete state vector obtained by the decoder D(·) from the compressed data z (S) Recovered;

[0168] h (S) (t): represents the altitude of the aircraft at time t;

[0169] represents the vertical acceleration of the aircraft at time t;

[0170] ψ (S) (t): represents the yaw angle of the aircraft at time t;

[0171] Unified decoder D(·) is a unified decoder with formal airworthiness approval and certification qualifications. It can restore the full timing state during the risk sampling phase and enforce the landing configuration constraints.

[0172] Inject the landing configuration constraint matrix C in real time during the decoding process cfg , the matrix coupling flap angle, touchdown attitude and tire pressure parameters, is forced to meet |ψ (s) (t)|≤ψ lim , a max is the maximum vertical acceleration threshold allowed during the landing phase, ψ lim is the maximum allowed yaw angle deviation to prevent latent space sampling from producing physically unreachable trajectories;

[0173] Among them, the landing configuration constraint matrix C cfg During the decoding phase, a linear mapping and boundary projection is performed on the decoder output variables such as altitude, vertical acceleration, yaw angle, flap position, and tire pressure. Its function is to force the trajectory parameters obtained by arbitrary sampling in the latent space to fall within the pre-defined legal landing configuration range, ensuring that the aircraft altitude does not drop below the ground, the vertical acceleration does not exceed the limit, the yaw angle is within the allowable range, and the certification standards for flap deployment and tire pressure are met, eliminating physically unattainable or illegal landing postures.

[0174] S54. Extract the hazard vector, scan each decoded trajectory one by one, and extract the peak vertical acceleration Window average drop rate and the maximum sliding angle T is the length of the time window for calculating the average decline rate, that is, the difference between the upper and lower limits of the integral, which is consistent with the prediction / backtracking window length defined above. The combination of the three is a single sample risk vector Complete characterization of impact strength, energy dissipation efficiency and lateral controllability;

[0175] S55. (s) Enter the hard landing risk function R(·);

[0176] If R(d (s) )≥1, it is judged as out of limit, and the indicator variable χ (s) =1, otherwise, χ (s) =0;

[0177] The overall risk probability estimate is defined as:

[0178]

[0179] Among them, z α / 2 is the standard normal quantile, is the strategy distribution entropy, λ is the entropy-variance coupling coefficient, and M is the number of latent samples of S52;

[0180] CI ± The confidence interval is used to quantify the fluctuation range of the risk probability estimate, including the standard error based on the binomial distribution and the strategy entropy compensation term;

[0181] The above formula takes into account both sampling variance and strategy uncertainty to compensate for the lack of confidence caused by the sparse prior of heavy landing events;

[0182] S56. Uncertainty quantification and confidence interval output, Forming upper and lower quantiles [P L ,P U ]. When the turbulence increases or the runway friction decreases, causing the strategy entropy to rise, CI ± Automatic relaxation to avoid outputting too narrow and misleading probability bands to pilots;

[0183] S57. Generating and persisting a level 3 alarm. The system determines based on thresholds τ1 and τ2:

[0184] If P U <τ1, output "normal", if P L ≥τ2, output "severe", otherwise output "mild";

[0185] Y (*) (t) The trajectory Y with the maximum risk function value (*) (s) is included in the warning message as the most dangerous sample, and the message is written into the visual engine and the black box channel;

[0186] Complete the four-level chain of potential action → physical trajectory → statistical risk → confidence warning, providing pilots and training instructors with real-time, quantitative and traceable hard landing risk indication;

[0187] S6. The visual system receives the warning message and the highest-risk trajectory, performs B-spline interpolation at 20ms intervals, and completes the body-to-world coordinate conversion. It maps the risk probability to color and transparency, then projects it into the volume shader. It also injects wind volume, runway friction map, and tire contact curve increment instructions into the physics engine. If the sum of the GPU and physics stepping times exceeds a threshold, the volume mesh resolution is automatically reduced to ensure real-time rendering.

[0188] Among them, S6 includes the following steps:

[0189] S61. Master clock in the visual system The risk level G∈{normal, mild, severe} and the confidence interval [P L ,P U ], V represents the master clock of the visual subsystem, and Q is the total number of frames of the visual master clock in one rendering / pipeline cycle;

[0190] The rendering scheduler reads the current frame number q now And the flight control main clock frame t0 and the rendering frame q now Establish a one-to-one mapping φ:t0→q now ;

[0191] This determines the benchmark time point for the touchdown-to-front wheel lift phase on the graphics pipeline to prevent phase drift between the subsequent physical scene and the graphics image.

[0192] S62. When G = severe, the rendering scheduler calls the highest risk trajectory passed in S5:

[0193] Y (*) (t)=[h (*) (t),x (*) (t),y (*) (t)];

[0194] Among them, Y (*) (t) represents the three-dimensional space position trajectory, h (*) (t) is the height, x (*) (t) is the x-axis position, y (*) (t) is the y-axis position;

[0195] First, five B-spline interpolations are performed at 20ms intervals, and then the aircraft heading angle Ψ and the graphics world coordinate system attitude matrix R are calculated. bw Complete the coordinate transformation:

[0196]

[0197] Where r0 is the absolute position vector of the runway threshold in the world system; -h (*) The negative sign makes the height measure along +z w Axially;

[0198] S63. Graphics pipeline based on the risk probability P(t) and uncertainty width ΔP(t) = P at each interpolation point U -P L Dynamic mapping of color and transparency;

[0199] The mapping function adopts a linear-exponential hybrid form:

[0200]

[0201] Here, β is fixed based on the display brightness LUT; the greater the uncertainty, the lower the transparency, avoiding misleading the pilot's attention at the risk boundary. All pixels are then projected into the volume shader via 3D texture coordinates, achieving a thermal cloud overlay that decreases with altitude.

[0202] S64. The rendering scheduler also parses Y (*) (t) Wind shear amplitude ΔW uploaded x , runway friction gap Δμ and tire slip margin λ s ;

[0203] These three physical indicators are discretized into a continuous interpolation sequence of 120 frames, and then converted into incremental instructions {δW f (q),δμ(q),δλ(q)}, ensuring that the wind field body, friction coefficient map and tire contact model change synchronously on the engine side;

[0204] S65. To maintain visual-force feedback consistency, the physics engine applies incremental commands frame by frame over the next 2 seconds of gameplay in the order of wind first, surface second, and wheel last.

[0205] Loading δW f (q) Update the six-degree-of-freedom wind field, then write δμ(q) to the runway material parameter table, and finally call the ground module interface to write δλ(q) into the tire longitudinal force-displacement curve;

[0206] Three-stage progressive processing avoids visual tearing or body jitter caused by instantaneous parameter jumps;

[0207] S66. At the end of each rendering frame, the scheduler reads the GPU rendering time T r (q) and physical step time T p (q), if T r (q)+T p(q) ≥ 30ms, immediately triggering resolution downsampling, reducing the length, width, and height of the volume grid by 25% each; at the same time, increasing the asynchronous shading priority to ensure that the visual system maintains a real-time response of ≥ 33Hz, meeting the human factors needs of pilots;

[0208] S67. When the two-second transition phase is complete, the rendering system writes the current resolution of the thermal grid, the final values ​​of the three parameters in the physics engine, and the time sequence of the entire update period to the log file LOG VR ;

[0209] The log is then aligned with the black box flight data via timestamps, allowing for subsequent playback of the entire impact-buffer-taxi process and providing labels for the next round of incremental learning, enabling continuous collaboration among the three domains of data, model, and vision within a closed training loop.

[0210] S7. After the aircraft speed drops below a preset threshold, the touchdown frame is locked as the synchronization origin. The aligned ground truth data packets are extracted and written into the incremental learning buffer. The vertical impact error is calculated, hard samples are selected, mixed with an equal number of random samples, and replayed into the policy network. Rapid adaptation is performed based on the weighted gradient of the hard samples. Simultaneously, the terminal encoding layer of the variational network is unfrozen and the dual-channel attention is fine-tuned using limited iterations. The energy conservation difference and damping ratio error are monitored in real time. If they exceed the threshold, the weights are immediately rolled back. After batch replay verification is completed in the isolated sandbox, the new weights are hot-swapped online.

[0211] Among them, S7 includes the following steps:

[0212] S71. Landing data collection is aligned with the time axis. After the aircraft has completely decelerated to below 30 kt, the flight control recording system immediately locks the touchdown frame number t * As the synchronization origin, continuously extract the IMU three-axis acceleration a imu (t), main wheel load L w (t), Airport LIDAR wind field profile W x,y,z (t, z) and the measured value of runway slipperiness μ w (t); All observations are mapped to the prediction time index by linear interpolation Generate alignment data packets Then write to the incremental learning buffer pool BUF IL .in, represents the sampling time of the full temporal prediction trajectory output by the unified decoder D(·) in step S5 within the prediction window;

[0213] S72. Hard sample calibration, for each record r in the buffer pool j Calculate the vertical impact error:

[0214]

[0215] in, From Synchronous prediction for S5, is the jth hard sample record, which is the true vertical acceleration obtained by linear interpolation from the IMU three-axis acceleration and other observations. The predicted trajectory of the j-th sample output by the S5 decoder at the same time The vertical acceleration generated. j Sort in descending order, select the top 10% samples and assign hard sample labels to highlight extreme landing scenes such as rain and snow runways and sharp shears, N P is the total number of prediction time indices, i.e. k = 1…N P ;

[0216] S73. Balanced distribution sampling: randomly extract common samples with the same number of hard samples from the buffer pool, and merge the two into a small batch sample set at a 1:1 ratio.

[0217] Prevent the network from overfitting the tail distribution and maintain the generalization balance of the strategy between normal and abnormal conditions;

[0218] S74. Rapid adaptation of the game strategy network to batch Replay to the input of the strategy network and use the real landing reward R i =-αε i +βμ w (t * ) re-evaluates the value of each sample. α and β are two hyperparameter weights used to balance the ratio of prediction error penalty and runway friction reward in the actual landing reward. Among them, the network parameters are based on weighted gradient:

[0219]

[0220] Among them, η is the adaptive learning rate, π θ is the current strategy, ω i is the normalized weight, which is used to amplify the contribution of hard samples to the gradient and achieve high-speed convergence to the latest runway friction conditions. i For small batches The "environmental state" input corresponding to the i-th sample, It is composed of the hard sample set and the common sample set fused at a 1:1 ratio in the previous step S73;

[0221] In S74, ω i The definition and function are as follows:

[0222]

[0223] Among them, ∈ i is the landing error (hard sample error) of the i-th playback sample, is the sample set;

[0224] Normalized weight ω i The error of each sample ∈ i Normalize across the entire batch so that Since the sample with large error ∈ i Larger, its corresponding ω i It is also larger and accounts for a higher proportion in the gradient update, allowing the policy network to converge quickly on hard samples first;

[0225] S75. Limited Fine-tuning of Variational Networks

[0226] Synchronize Input variational network, only unfreeze the terminal encoding layer and the dual-channel attention of the decoding side, and lock the rest of the weights, limiting E max = 50 iterations, completing gradient updates in independent GPU streams to ensure that the entire training-inference pipeline has no mutual exclusion waits;

[0227] S76. Physical constraint monitoring and weight rollback, real-time evaluation of energy conservation differences during fine-tuning and the damping ratio error Δζ, t represents the real-time sampling moment, Substituting the velocity obtained by observation interpolation into the kinetic energy formula, we can get: The velocity calculated from the decoder output trajectory is substituted into the kinetic energy formula;

[0228] like Or if |Δζ| exceeds the certification threshold, it will immediately roll back to the last stable checkpoint to prevent the model from overfitting to the small batch and destroying the physical consistency;

[0229] S77. Sandbox verification and security switching. After training, the system executes 10 4 Play back historical landing samples and calculate the average inference time and the error quantile ε 95％ .like And ε 95％ If it is not increased, it is written into the online weight table and hot-switch the version pointer to provide immediate incremental improvement for the next landing call, while retaining the old version for failure fallback.

[0230] The present invention provides a deep learning-based short-term abnormal risk warning and visual simulation adaptive method for the aircraft landing phase. This method addresses the shortcomings of existing landing risk warning systems, such as reliance on empirical thresholds, neglect of short-term dynamic changes, and disconnection between simulation scenarios and real-time risks. The present invention constructs a latent variable space based on deep learning, fully covers the tail of the action with weighted random sampling, decodes and restores the full temporal trajectory, extracts risk vectors such as vertical acceleration, descent rate, and yaw angle slip, and quantifies uncertainty with credible intervals to achieve real-time and accurate short-term abnormal risk warning. Simultaneously, the adaptive visual simulation is driven by a real-time risk index, seamlessly coupling ground friction, scene changes, and visual rendering, improving the fidelity of training scenarios and enhancing the rapid online learning capabilities of the policy network. This method effectively reduces missed and false alarm rates, improves landing safety, and enhances pilot training efficiency, providing a complete solution for intelligent aircraft landing risk management and traceable simulation.

[0231] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems or computer program products. Therefore, the application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal translation scripting language JavaScript, etc.

[0232] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0233] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1The function specified in one or more boxes.

[0234] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0235] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0236] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A method for short-term abnormal risk warning and visual simulation adaptiveness during aircraft landing phase based on deep learning, characterized in that: include: S1. Acquire four raw signals: the inertial measurement unit, radio altimeter, fuselage longitudinal accelerometer, and airborne radar echo. After compensating for the pre-calibrated propagation delay, map each signal to a unified time grid using a cubic spline kernel function. Switch to landing filter parameters at the decision altitude, define a high-frequency impact window based on the gear resonance template to extract the impact component, use a sliding interval operation to obtain the wind shear component, and obtain the residual component through orthogonal projection of the ground effect energy. Write the three-channel signals (impact, wind shear, and residual) along with the timestamp into a ring buffer. S2. Receive data from the ring buffer on the flight control main time axis, backtrack a preset number of seconds to form a landing window, map impact, wind shear, and residuals into aircraft node features, wind field node features, and runway node features, and inject real-time runway friction coefficient and wind shear correction weights. Then, extract spatiotemporal correlation features through causal dilation convolution and multi-scale graph convolution, concatenate them into a state vector, and input it into the policy network. The policy network outputs a two-speed maneuver sequence and its probabilities. In the event of vertical kinetic energy exceeding the limit, a negative reward is used to trigger an online update using Adam-SGD with learning rate gate. S3. Query the current aircraft model's ground effect function and gear compression curve to form an expected energy attenuation template. Perform orthogonal projection on the convolution-recurrent features to obtain conserved features, which are then mapped to a latent vector. Hadamard weighting is applied to the latent vector based on the policy probability, and the resulting vector is input into a dual-channel attention decoder to obtain the predicted vertical acceleration, sink rate, and wheel load trajectory for the next 2 seconds. The reconstruction error is then calculated by comparing the predicted trajectory with the measured data. S4. Generate a soft index gating coefficient based on the real-time runway friction coefficient and wind farm risk score. Apply the gating coefficient to the latent vector to obtain a weighted latent vector. Use the decoder to quickly infer the weighted latent vector and compare it with the predicted trajectory in S3. Map a real-time penalty based on the error and write it back to the policy network. S5. Perform Gaussian perturbation sampling within the weighted latent space according to the modified strategy probability. Restore the full time-series trajectory from the sampled latent vectors through a unified decoder. Extract the peak vertical acceleration, window average sink rate, and maximum yaw angle. Calculate the hard landing risk probability and output a confidence interval based on the sampling variance and strategy entropy. Generate a normal, mild, or severe warning message. S6. The visual system receives the warning message and the highest-risk trajectory, performs B-spline interpolation at intervals, and completes the body-to-world coordinate conversion. It maps the risk probability to color and transparency, then projects it into the volume shader. It also injects wind volume, runway friction map, and tire contact curve increment instructions into the physics engine. If the sum of the GPU and physics stepping times exceeds a threshold, the volume mesh resolution is automatically reduced. S7. After the aircraft speed drops below a preset threshold, the touchdown frame is locked as the synchronization origin. The aligned true value data packets are extracted and written into the incremental learning buffer. The vertical impact error is calculated, hard samples are selected, mixed with an equal number of random samples, and replayed into the policy network. Fast adaptation is performed based on the weighted gradient of the hard samples. Simultaneously, the terminal encoding layer of the variational network is unfrozen and the dual-channel attention is fine-tuned with limited iterations. Monitor the energy conservation difference and damping ratio error in real time. If the threshold is exceeded, roll back the weight immediately, and hot-switch the new weight online after batch replay verification is completed in the isolated sandbox.

2. The deep learning-based short-term abnormal risk warning and visual simulation adaptive method for aircraft landing phase according to claim 1 is characterized by: S1 includes the following steps: S11. Timestamp synchronization and unified resampling, establishing a unified time grid with a length of Δt = 1ms {t k |k=1,…,K}; Assume that the four original time stamp sequences of the inertial measurement unit IMU, radio altimeter, fuselage longitudinal accelerometer and airborne radar echo are T I ,T H ,T B ,T R , the corresponding observation vector is x I,p ,x H,p, ,x B,p ,x R,p , where the subscript p represents the respective original sampling points; The fixed propagation delay compensation pre-calibrated for each sensor is denoted as τ I , τ H ,τ B ,τ R ; Perform cubic spline kernel interpolation mapping on the uniform time grid to obtain the aligned observation vector y(t k ), the mapping relationship satisfies: Among them, s∈{1,2,3,4}, corresponding to IMU, radio altimeter, fuselage longitudinal accelerometer, and airborne radar echo in sequence; N s is the number of original samples in the s-th path; κ s (·) is the cubic spline kernel function designed for the dynamic bandwidth of the road; e s is a 4×1 basis vector; x s,p is the observation value of the s-th sensor at the p-th original sampling point; τ s Fixed propagation delay compensation for each channel; S12. Decision height detection and dedicated filter channel switching, in real-time stream y(t k ) to extract the radio height H(k)=y H (t k ) and vertical descent rate V z (k) = (H(k) - H(k-1)) / Δt; Once it is detected that H(k)≤H D , immediately turn off the high-altitude filter used in the cruise phase, start the landing-specific complementary filter channel, and load the ground effect reference model g(k) from the aircraft type-mass-flap angle three-dimensional index table, where H D Decision altitude thresholds for combined aircraft type-runway certification; S13. Gear resonance matching and shock high-frequency window calibration, for the aligned observation vector y of the fuselage longitudinal acceleration B (t k ) online calculation of its short-time Fourier transform power spectrum P(k,f), and retrieval of the resonance template Ψ(f) under the current tire pressure and load conditions from the landing gear mechanical feature library; Solve the cross-correlation C(t k ), and then C(t k ) back-calculated to the time domain high-frequency impact window w c , in w c Only keep y B (t k ) high frequency component, and obtain the calibrated shock pulse i c (k); Perform short-time Fourier transform (STFT) on the signal of each sliding window length L1: P(k,f)=STFT{y B (t k )},P(k,f)=|A(k,f)| 2 ; Where P(k,f) is the short-time Fourier transform power spectrum; According to the current tire pressure and load conditions, the frequency domain template Ψ(f) is taken from the gear mechanical feature library and the frequency domain template Ψ(f) is taken from the gear mechanical feature library. k Calculate the cross-correlation between the power spectrum and the template: The upper and lower limits of the integral are (f min ,f max ), peak point Corresponding to the central moment of a landing impact; The high-frequency impact window is defined using the empirical window width Δt and the template bandwidth Δf as parameters: w c =[t d -Δt,t d +Δt]; And perform bandpass filtering of the center frequency and bandwidth Δf on the original acceleration within this time domain area: i c (t)=BandPass{y B (t),f ± Δf},t∈w c ; The above time domain impulse signal i c (t) According to the unified time grid {t k }Interpolate or resample to obtain a discrete sequence: i c (k)=i c (t k ); S14. Low-frequency wind shear feature separation, for the area outside the window Ω={t k }\W c , parallel calculation of IMU lateral acceleration within the sliding interval of length M The velocity difference ΔV(t k )=y R (t k )-y R (t k-1 ), the index is obtained by offsetting the antenna-INS installation error based on the correction coefficient β If |Λ(k)|≥σ Λ , then mark the wind shear event in Ω and generate the wind shear feature sequence c w (k), where σ Λ is the threshold; S15. Based on the energy conservation principle, the ground effect residual denoising is performed. The ground effect reference model g(k) is injected into the observation matrix of the Kalman-complementary fusion device, and the longitudinal acceleration a(k)=y is calculated within the current analysis window length M. B (t k ) Perform energy orthogonal projection and calculate the residual: Where the numerator is the inner product of a(k) and g(k), the denominator is the energy of g(k) within the window, and the fraction as a whole is the optimal scaling factor in the least squares sense; S16. The shock pulse i calibrated by S13 c (k) Wind shear characteristic sequence c separated by S14 w (k) is concatenated with the residual R(k) obtained by S15 in ascending order of time index to form a three-channel fusion signal F(k) = [i c (k),c w (k), R(k)], F(k) is written to the shared address BUF1 through the double-ended ring buffer and appended with the complete timestamp table t k .

3. The method for short-term abnormal risk warning and visual simulation adaptively during aircraft landing phase based on deep learning according to claim 2 is characterized by: S1 includes the following steps: S21. Receive and align the three-channel fusion sequence, on the flight control main axis The three-channel fusion signal F(t k )=[i c (t k ),c w (t k ),R(t k )], K in the flight control main time axis represents the total number of sampling points on the main time axis, and F is the three-channel fusion signal function in S16; F(t k ) is mapped to the main time axis through linear interpolation to obtain the alignment result The current interpretation time t now Backtrack L seconds for the right endpoint to build a continuous landing window Where N = L / Δt, k0 is the current judgment time t now The index in the timestamp sequence; S22. Multi-mode dynamics-environment initial feature map, in the window Internal pair Channel analysis, using the aircraft airspeed meter to read the mass m a ; The vertical velocity is obtained by first-order difference Wheel load obtained from main wheel strain gauge Obtaining the crosswind component from the inertial navigation lateral velocity and the meteorological blade meter data The three observations are mapped into time series features of aircraft nodes, runway nodes, and wind field nodes. Each feature is scaled according to the aircraft model normalization scaling coefficient vector e = [s1,…,s5] to construct the node tensor: The symbols are explained as follows: X u,j,l : the eigenvalue of the u-th node, time j, and l-th dimension in the graph network; u∈{1,2,3}: node type index, where 1 = aircraft, 2 = runway, 3 = wind farm; j=1,…,N: time index, N is the total number of frames in the lookback window; l: Feature dimension number, corresponding to the five observations one by one, and other positions are filled with 0; e = [s1, s2, s3, s4, s5]: normalized scaling coefficient vector for dimension matching; The vertical velocity of the flight control clock at frame j; The main wheel load at frame j; μ w : real-time runway slipperiness coefficient; P t : Real-time tire inflation pressure; Crosswind component of the jth frame; The jth frame time after the flight control main clock is aligned; S23. Dynamic weight coupling of sparse environmental attributes, based on the μ uploaded in real time by the apron weather radar and runway friction measurement device. w With P t , the second and third columns of the runway node feature matrix are written immediately, and then passed through the weight function W p (k)=α1μ w +α2 / P t Generate shock weight modifiers; S24. Temporal convolution - multi-scale graph convolution joint state encoding, input the aircraft node along the time dimension into the causal expansion convolution network to obtain the time feature h T (k); The runway nodes and wind farm nodes are fed into a three-layer multi-scale graph convolution together with the dynamic adjacency matrix A(k) constructed by S23, and the spatial correlation features h are extracted using a two-order approximation. G (k); In the fusion layer, h T (k) and h G (k) performs channel-level splicing and outputs the complete game state vector s(k) = [h T (k)‖h G (k)]; The dynamic adjacency matrix A(k) in S23 is composed of the basic topology and the dynamic weights of the environmental attributes. Assuming that the three types of nodes are 1 = aircraft, 2 = runway, and 3 = wind farm, the basic adjacency matrix is ​​as follows: Among them, the diagonal ring is set to 1, The rest are 0; The dynamic weight increment is as follows: Δa 12 (k)=Δa 21 (k)=g r (μ w (k)),Δa 23 (k)=Δa 32 (k)=g w (φ w (k)), The symbols are explained as follows: μ w (k): runway friction coefficient at time k; φ w (k): wind farm risk score at time k; g r ,g w : A weight mapping function that can be learned or empirically fitted; The dynamic adjacency matrix is ​​as follows: A(k)=A 0 +ΔA(k); That is: Other positions remain disconnected or self-loop constants; S25. Hard landing event detection and online reward update, calculating vertical kinetic energy based on the aircraft inertia parameters at each time step Assume the maximum allowable kinetic energy threshold E lim , if E k (k)≥E lim , it is considered a hard landing and rewards are generated immediately: Among them, λ>0 is the scale coefficient, η>1 is used to perform high-order amplification of the excess kinetic energy, and ρ is the slipperiness gain; S26. Policy network adaptive evolution and maneuver sequence generation. The policy network receives the previous parameters and the new reward r(k). Within the same computational graph, it completes one iteration using an Adam-SGD hybrid optimization with learning rate gating. The updated network generates two maneuver action sequences under the current state s(k): a short-term sequence π1 of length T1 = 3s and a medium-term sequence π2 of length T2 = 5s, and outputs the corresponding probabilities p1(k) and p2(k). S27. The risk assessment head outputs a priori risk set, and each candidate sequence π i Through rapid simulation of the aircraft-runway-wind field hybrid dynamics model, the vertical kinetic energy peak is calculated separately. and residual friction margin Constructing risk scores based on dual indicators Among them, w1 and w2 are experience weights, which eventually form the risk score set {ρ1, ρ2} and the strategy probability {p1(k), p2(k)} which are output in parallel.

4. The method for short-term abnormal risk warning and visual simulation adaptively during aircraft landing phase based on deep learning according to claim 3 is characterized by: S3 includes the following steps: S31. Align and crop the prediction window. The flight control computer advances with a 1000 Hz master clock at the end of landing. When the real time corresponding to the current frame number is t now When it is determined to be in the radar altitude range of 2-3m before the aircraft touches down, the system synchronously receives the three-channel fusion signal F(t k )=[i c (t k ),c w (t k ),R(t k )] and the strategy probability p(k) output by S2, first k ) and the main time axis Do millisecond interpolation alignment, and then use T P = 3s window W to intercept data backward, T p Specifically refers to the length of the forecast / lookback window; S32. Transient-gradient dual feature extraction, in the window Internal construction tensor U k , whose first row i c (t k ) reflects the main wheel rigid impact peak, the second row c w (t k ) describes the low-frequency lateral shear driven by the crosswind shear, and the third row R(t k ) records sensor residual noise. For the spectral differences of impact (short pulses on the order of milliseconds) and wind shear (slowly varying on the order of hundreds of milliseconds), a one-dimensional convolution is first performed with a learnable convolution kernel W1 of length 7 to enhance the high-frequency components and highlight the narrow peak caused by gear pressure drop. The convolution result is then progressively input into a gated recurrent unit, which uses a recurrent gate to accumulate the crosswind slope over a time constant of 250-500ms. Among them, the convolution-recurrence coupling satisfies: H k =GRU(σ(W1*U k +b1),H k-1 ); Where σ = ELU, the physical property of symmetrical stretching corresponding to the impact amplitude, and b1 is the bias; S33. Build the expected energy attenuation template and obtain the current angle of attack A α , lower washing angle A ω The dimensionless ground effect function g(k,A α ,A ω ); Then retrieve the gear compression curve from the tire pressure-oil pressure-temperature triplet and calculate the instantaneous stiffness-damping coefficient C st ; The two parts are combined into the expected energy decay template: T(k)=C st ·g(k,A α ,A ω ); S34. Physical constraint suppression and latent space consistency, if the convolution-recurrent feature Z(k) = H k At a certain moment, the energy exceeds the template, and the excess energy is eliminated through orthogonal projection to obtain the conserved feature Z sup (k); S35. Inject the strategy probability conditional latent vector into the strategy probability p(k). If p1(k) is high, it means that the system predicts that the pilot is biased towards rapid correction; if p2(k) is dominant, it means that the pilot accepts a longer buffer. S36. Dual-channel attention decoding prediction trajectory, the decoder first constructs the impact channel attention Att for z(k) a , whose attention is drawn from the gear pressure drop reference signal, specifically capturing the strong nonlinear peak of <20ms; Secondly, construct the aerodynamic slowly varying channel attention Att b , its attention is based on the ground effect air cushion attenuation time constant, focusing on the 200-500ms slip trend, the two-channel output is gated and fused to generate the vertical acceleration, sinking rate, and original predicted trajectory for the next 2 seconds S37. Reconstruct the error and the posterior output, and decode the original prediction trajectory Measured with the window Do the two-norm difference to obtain the reconstruction error ε(k), and calculate the posterior mean μ of the latent variable in parallel z (k) and variance Will Write to shared buffer BUF2.

5. The method for short-term abnormal risk warning and visual simulation adaptively during aircraft landing phase based on deep learning according to claim 4 is characterized by: S4 includes the following steps: S41. Time alignment of policy probability and latent variable and buffer writing. 500ms before the aircraft's main wheels touch down, read the policy probability sequence p(k) = [p1(k), p2(k)] generated by S2 and the posterior latent variable z(k) output by S3. First, align p(k) to the latent variable timestamp by linear interpolation Then follow the index Write to the shared buffer BUF3, where p in the index represents the latent variable and n represents the nth one; S42. Hazard preference gate coefficient generation: real-time risk score ρ given by wind farm LiDAR and black surface friction vehicle at 500m from the runway end n (k) and the strategy probability p(k) jointly determine the degree of dynamic stress, and the gating coefficient is constructed based on the soft index mapping: Among them, κ is calibrated online based on the current tire-track friction coefficient. w decreases, κ automatically increases, and n represents the signal of the nth gated channel; S43. Gated modulation forms a weighted latent space, multiplying g(k) element-wise by the latent variable z(k) to generate a weighted latent vector S44. Weighted latent space decoding and reconstruction error evaluation, Input the fast decoder with the same structure as step S3 but frozen parameters to obtain the weighted prediction trajectory and with S3 original prediction Do point-by-point alignment; S45. Error-penalty mapping and writing back to the policy network, mapping the reconstructed error vector bi-norm into real-time penalty: Among them, λ is adaptively adjusted based on the aircraft model's decision altitude minus the remaining runway length. High-error and high-gating actions will receive the largest negative reward, which is directly written back to the corresponding branch of the policy network. S46. Quickly fine-tune and update the policy probability in the same graph. The policy network uses Δr in the same computation graph. n (k) Perform a restricted-step Adam-SGD update; Generate revised strategy probability immediately after update To reflect the latest decision-making tendency of the three-party collaboration of pilot, autopilot and risk gate; S47. The system will correct the probability Weighted latent vector and gated risk score The message is framed into a policy latent space-risk ternary message and written into the shared buffer BUF4.

6. The method for short-term abnormal risk warning and visual simulation adaptively during aircraft landing phase based on deep learning according to claim 5 is characterized by: S5 includes the following steps: S51. When the flight control main clock detects that the main wheel pressure drop curve has crossed the 80% touchdown threshold, it immediately reads the corrected strategy probability output by S4 With weighted latent variables S52. Probability weighted latent space lottery, perform weighted lottery according to the modified strategy probability: each time from the discrete distribution Select action number n∈{3s,5s} from the corresponding latent variable subset Gaussian perturbation is applied to Generate sample z (s) , the weight-perturbation joint sampling satisfies: Where M is the number of samples, σ z Automatically set based on the S3 posterior variance to ensure that the proportion of high-risk actions in the sample pool is not less than 70%. is the diagonal covariance matrix, I is the identity matrix, z (s) =v means the sth latent sample takes value v; S53. (S) Input the airworthiness-approved unified decoder D(·) to restore the full time sequence trajectory. The formula is as follows: Among them, Y (S) (t) is a more complete state vector obtained by the decoder D(·) from the compressed data z (S) Recovered; h (S) (t) represents the altitude of the aircraft at time t; represents the vertical acceleration of the aircraft at time t; ψ (S) (t) represents the yaw angle of the aircraft at time t; Inject the landing configuration constraint matrix C in real time during the decoding process cfg , the matrix coupling flap angle, touchdown attitude and tire pressure parameters, is forced to meet a max is the maximum vertical acceleration threshold allowed during the landing phase, ψ lim is the maximum allowable yaw angle deviation; S54. Extract the hazard vector, scan each decoded trajectory one by one, and extract the peak vertical acceleration Window average drop rate and the maximum sliding angle T is the length of the time window for calculating the average decline rate, and the three are combined to form a single sample risk vector. Complete characterization of impact strength, energy dissipation efficiency, and lateral controllability; S55. (s) Enter the hard landing risk function R(·); If R(d (s) )≥1, it is judged as out of limit, and the indicator variable χ (s) =1, otherwise, χ (s) =0; The overall risk probability estimate is defined as: Among them, z α / 2 is the standard normal quantile, is the strategy distribution entropy, λ is the entropy-variance coupling coefficient, and M is the number of latent samples of S52; CI ± The confidence interval is used to quantify the fluctuation range of the risk probability estimate, including the standard error based on the binomial distribution and the strategy entropy compensation term; S56. Uncertainty quantification and confidence interval output, Forming upper and lower quantiles [P L ,P U ], when turbulence increases or runway friction attenuates, causing the strategy entropy to rise, CI ± Automatic relaxation; S57. Generating and persisting a level 3 alarm. The system determines based on thresholds τ1 and τ2: If P U <τ1, output "normal", if P L ≥τ2, output "severe", otherwise output "mild"; Y (*) (t) The trajectory Y with the maximum risk function value (*) (s) is included in the warning message as the most dangerous sample, and the message is written into the visual engine and the black box channel.

7. The method for short-term abnormal risk warning and visual simulation adaptively during aircraft landing phase based on deep learning according to claim 6 is characterized by: S6 includes the following steps: S61. Master clock in the visual system The risk level G∈{normal, mild, severe} and the confidence interval [P L ,P U ], V represents the master clock of the visual subsystem, and Q is the total number of frames of the visual master clock in one rendering / pipeline cycle; The rendering scheduler reads the current frame number q now And the flight control main clock frame t0 and the rendering frame q now Establish a one-to-one mapping φ:t0→q now ; S62. When G = severe, the rendering scheduler calls the highest risk trajectory passed in S5: Y (*) (t)=[h (*) (t),x (*) (t),y (*) (t)]; Among them, Y (*) (t) represents the three-dimensional space position trajectory, h (*) (t) is the height, x (*) (t) is the x-axis position, y (*) (t) is the y-axis position; Perform five B-spline interpolations at 20ms intervals, and then calculate the attitude matrix R based on the aircraft heading angle Ψ and the graphics world coordinate system attitude matrix R. bw Complete the coordinate transformation: Where r0 is the absolute position vector of the runway threshold in the world system; -h (*) The negative sign makes the height measure along +z w Axially; S63. Graphics pipeline based on the risk probability P(t) and uncertainty width ΔP(t) = P at each interpolation point U -P L Dynamic mapping of color and transparency; The mapping function adopts a linear-exponential hybrid form: Where β is fixed according to the display brightness LUT; the greater the uncertainty, the lower the transparency; S64. The rendering scheduler also parses Y (*) (t) Wind shear amplitude ΔW uploaded x , runway friction gap Δμ and tire slip margin λ s ; These three physical indicators are discretized into a continuous interpolation sequence of 120 frames, and then converted into incremental instructions {δW f (q),δμ(q),vλ(q)}; S65. The physics engine applies incremental instructions frame by frame over the next 2 seconds of game time, in the order of wind first, surface second, and rear wheel. Loading δW f (q) Update the six-degree-of-freedom wind field, then write δμ(q) to the runway material parameter table, and finally call the ground module interface to write δλ(q) into the tire longitudinal force-displacement curve; S66. At the end of each rendering frame, the scheduler reads the GPU rendering time T r (q) and physical step time T p (q), if T r (q)+T p (q) ≥ 30ms, immediately trigger resolution downsampling, reducing the length, width and height of the volume grid by 25% each; at the same time, increase the asynchronous shading priority to ensure that the visual system maintains a real-time response of ≥ 33Hz; S67. When the two-second transition phase is complete, the rendering system writes the current resolution of the thermal grid, the final values ​​of the three parameters in the physics engine, and the time sequence of the entire update period to the log file LOG VR .

8. The method for short-term abnormal risk warning and visual simulation adaptively during aircraft landing phase based on deep learning according to claim 7 is characterized by: S7 includes the following steps: S71. Landing data collection is aligned with the time axis. After the aircraft has completely decelerated to below 30 kt, the flight control recording system immediately locks the touchdown frame number t * As the synchronization origin, continuously extract the IMU three-axis acceleration a imu (t), main wheel load L w (t), Airport LIDAR wind field profile W x,y,z (t, z) and the measured value of runway slipperiness μ w (t); All observations are mapped to the prediction time index by linear interpolation Generate alignment data packets Then write to the incremental learning buffer pool BUF IL ,in, is the sampling time of the full temporal prediction trajectory output by the unified decoder D(·) in S5 within the prediction window; S72. Hard sample calibration, for each record r in the buffer pool j Calculate the vertical impact error: in, From a imu , Synchronous prediction for S5, is the jth hard sample record, the true vertical acceleration obtained by linear interpolation of the IMU three-axis acceleration, The predicted trajectory of the j-th sample output by the S5 decoder at the same time The generated vertical acceleration is ε j Sort in descending order, select the top 10% samples and assign hard sample labels, N p is the total number of prediction time indices, i.e. k = 1…N p ; S73. Balanced sampling: randomly extract common samples with the same number of hard samples from the buffer pool, and merge them in a 1:1 ratio to form a small batch sample set B = {r h,i }∪{r n,i }; S74. Rapid adaptation of the game strategy network to batch Replay to the input of the strategy network and use the real landing reward R i =-αε i +βμ w (t * ) re-evaluates the value of each sample. α and β are two hyperparameter weights used to balance the ratio of prediction error penalty and runway friction reward in the actual landing reward. The network parameters are based on weighted gradients: Among them, η is the adaptive learning rate, π θ is the current strategy, ω i is the normalized weight, s i For small batches The "environmental state" input corresponding to the i-th sample, It is composed of the hard sample set and the common sample set fused at a 1:1 ratio in the previous step S73; ω i is defined as follows: Among them, ∈ i is the landing error of the i-th playback sample, is the sample set; Normalized weight ω i The error of each sample ∈ i Normalize across the entire batch so that S75. Limited Fine-tuning of Variational Networks Synchronize Input variational network, only unfreeze the terminal encoding layer and the dual-channel attention of the decoding side, and lock the rest of the weights, limiting E max = 50 iterations, completing gradient updates in independent GPU streams to ensure that the entire training-inference pipeline has no mutual exclusion waits; S76. Physical constraint monitoring and weight rollback, real-time evaluation of energy conservation differences during fine-tuning and damping ratio error t represents the real-time sampling time, Substituting the velocity obtained by observation interpolation into the kinetic energy formula, we can get: The velocity calculated from the decoder output trajectory is substituted into the kinetic energy formula; like or If the authentication threshold is exceeded, it will immediately roll back to the last stable checkpoint; S77. Sandbox verification and security switching. After training is completed, the system executes 10 4 Play back historical landing samples and calculate the average inference time and the error quantile ε 95％ ,like And ε 95％ Without increasing the value, the value is written to the online weight table and hot-switch the version pointer to provide immediate incremental improvement for the next landing call, while retaining the old version for failure fallback.

Images