Flight altitude prediction method based on neural network

Abstract

The present invention is applicable to the field of flight altitude prediction technology, and in particular relates to a flight altitude prediction method based on a neural network. The method comprises: collecting sensor data from a passenger aircraft, suppressing outlier data points through a dynamic truncation mechanism based on a distribution quantile threshold; partitioning the sensor data using a time series slicing strategy to obtain a training set and a validation set, and constructing a prediction feature set based on the training set and the validation set; constructing a flight altitude prediction model, training the flight altitude prediction model using the prediction feature set, and predicting flight altitude based on the trained flight altitude prediction model. The present invention applies normalization technology to perform data reduction on flight time series signals, designs a parallel network structure including a multi-scale convolution kernel feature extraction layer and an attention weight allocation module, and performs gradient iteration on the Huber loss function through an optimizer that dynamically adjusts the learning rate, significantly improving the model's ability to characterize nonlinear change patterns in flight states.

Description

Technical Field

[0001] The present invention belongs to the technical field of flight altitude prediction, and in particular relates to a flight altitude prediction method based on a neural network. Background Art

[0002] As air transport continues to expand, flight altitude prediction plays an increasingly crucial role in safety management and fuel economy optimization. While CNNs and LSTMs demonstrate advantages in modeling spatiotemporal features, existing approaches face two key challenges: insufficient modeling of the physical relationships between multi-source, heterogeneous data, resulting in a lack of contribution from key flight dynamics features; and weak mechanisms for suppressing sudden noise interference, limiting the model's robustness under complex operating conditions. These bottlenecks make it difficult for traditional prediction frameworks to meet the engineering requirements for high-precision aviation safety monitoring.

[0003] Although the CNN-LSTM joint architecture adopted by existing technologies realizes the parallel extraction of spatiotemporal features, when processing heterogeneous information streams such as meteorological observation parameters and construction time series data, it may lead to the omission of key spatial local patterns due to the lack of a hierarchical fusion mechanism. Summary of the Invention

[0004] The purpose of the present invention is to provide a flight altitude prediction method based on a neural network, aiming to solve the problem that when the existing technology processes heterogeneous information streams such as meteorological observation parameters and construction time series data, the key spatial local patterns may be missed due to the lack of a hierarchical fusion mechanism.

[0005] The present invention is achieved by providing a method for predicting flight altitude based on a neural network, the method comprising:

[0006] Collect sensor data from passenger aircraft and suppress outlier data points through a dynamic truncation mechanism based on distribution quantile thresholds;

[0007] The sensor data is divided into training set and validation set using time series slicing strategy, and the prediction feature set is constructed based on the training set and validation set.

[0008] Construct a flight altitude prediction model, train the flight altitude prediction model through the prediction feature set, and predict the flight altitude based on the trained flight altitude prediction model.

[0009] Preferably, the sensor data includes at least wind direction, wind speed, ground speed and angle of attack.

[0010] Preferably, in the step of suppressing outlier data points by a dynamic truncation mechanism based on a distribution quantile threshold, the dynamic truncation mechanism based on a distribution quantile threshold is expressed as:

[0011] ;

[0012] in, Indicates the height data after cleaning, Represents the original height data, and They represent the 2% and 98% quantiles of the height parameter, respectively.

[0013] Preferably, the step of collecting sensor data of the passenger aircraft and suppressing outlier data points through a dynamic truncation mechanism based on a distribution quantile threshold further includes normalizing the sensor data, which is expressed as:

[0014] ;

[0015] in, represents the standardized data, represents the original data points, represents the median of the data set X, represents the interquartile range of the data set X.

[0016] Preferably, the sensor data is divided by using a time series slicing partitioning strategy to obtain the training set and the validation set. The data sample ratio of the training set and the validation set is 9:1. When dividing the validation set and the training set, a stratified sampling mechanism is used to extract sample data.

[0017] Preferably, in the step of constructing a prediction feature set based on the training set and the validation set, the prediction feature set is constructed through three-stage feature engineering. First, the core physical features of the aircraft are used to construct enhanced features using the following formula:

[0018] ;

[0019] ;

[0020] ;

[0021] in, is the true airspeed of the aircraft, is the wind speed, It's ground speed. is the air density, is the lift coefficient, is the aircraft's angle of attack, is the indicated airspeed;

[0022] Time series dynamic characteristics are established. Time window statistics are performed on wind speed, ground speed, and airspeed parameters respectively. The instantaneous fluctuations of wind speed and ground speed are smoothed by the three-point moving average method. The instantaneous change rate of flight parameters is quantified using the first-order difference method, and the absolute value is converted into the change value. The oscillation amplitude of the flight parameters is measured using the five-point volatility method to characterize the stability of the aerodynamic environment as a risk warning feature. The formula is expressed as follows:

[0023] ;

[0024] ;

[0025] ;

[0026] in, is a 3-point moving average. For time point Parameter value, is the first-order difference of parameter x at time t, is the parameter value at time t, is the parameter value at time t-1, is the 5-point volatility at time t, For time Parameter value, is the parameter average value from time window t-4 to t;

[0027] The wind-to-ground speed ratio is introduced to characterize the aerodynamic environment, and the formula is expressed as:

[0028] ;

[0029] in, is the wind-to-ground speed ratio, is the wind speed, is the ground speed, Represents a very small constant.

[0030] Preferably, the flight altitude prediction model adopts a multimodal feature fusion framework, including a multimodal feature processing module. The multimodal feature processing module adopts a dual-channel processing architecture, including a time series feature extraction channel: a deep convolutional network is constructed to process time series features, and its mathematical expression is:

[0031] ;

[0032] ;

[0033] ;

[0034] in, Represents the first layer convolution output, using 32 groups of 5×1 convolution kernels to extract local fluctuation features. represents a one-dimensional convolution with a kernel size of 5, It is multi-channel time series data; is the sub-layer downsampling feature, Represents a one-dimensional convolution with a kernel size of 3; It is a compressed feature vector generated by adaptive average pooling; the activation function uses GELU; Maxpool selects the maximum value in the local window as the output to achieve nonlinear feature screening; AdaptiveAvgPool dynamically calculates the pooling window size to compress time series features of any length into a vector of fixed dimension;

[0035] Physical feature enhancement pathway: Construct a dynamic adjustment mechanism for attention weights. The calculation process is as follows:

[0036] ;

[0037] Where, represents the attention weight, σ represents the Sigmoid function, 、 is the learnable weight matrix, is the physical characteristic matrix.

[0038] Preferably, the features processed by the dual channels are fused by tensor concatenation, and the expression is:

[0039] ;

[0040] in, represents the fused feature vector, Represents the output of the temporal feature extraction channel, and ⊙ represents the Hadamard product operation.

[0041] Preferably, the fused feature vector is input into a deep regression network composed of a GELU activation function and a Dropout layer. The specific calculation formula is as follows:

[0042] ;

[0043] ;

[0044] ;

[0045] Where: 、 Corresponding to the primary regression layer output and the secondary regression layer output, respectively, 、 、 The weight matrices corresponding to the primary regression layer, secondary regression layer and output layer are respectively, where the predicted target It is a highly predictive value; Dropout is a random regularization technique that discards neuron connections with a certain probability during training; The GELU function adaptively gates the input features through a probability weighting mechanism.

[0046] Preferably, when collecting sensor data of a passenger aircraft, the sampling frequency is once per second.

[0047] The neural network-based flight altitude prediction method provided by the present invention applies normalization technology to perform data reduction on flight timing signals, designs a parallel network structure including a multi-scale convolution kernel feature extraction layer and an attention weight allocation module, and performs gradient iteration on the Huber loss function through an optimizer that dynamically adjusts the learning rate, significantly improving the model's ability to characterize the nonlinear change pattern of the flight state. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of a network structure provided by an embodiment of the present invention;

[0049] Figure 2 A comparison chart of prediction results corresponding to the QAR-1 dataset provided in an embodiment of the present invention;

[0050] Figure 3 A comparison chart of prediction results corresponding to the QAR-2 dataset provided by an embodiment of the present invention;

[0051] Figure 4 A comparison chart of prediction results corresponding to the QAR-3 dataset provided by an embodiment of the present invention;

[0052] Figure 5 A schematic diagram for comparing the results of the final approach phase provided by an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram for comparing results below 1000 feet provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0055] like Figure 1 As shown, an embodiment of the present invention provides a flight altitude prediction method based on a neural network, the method comprising:

[0056] Collect sensor data from passenger aircraft and suppress outlier data points through a dynamic truncation mechanism based on distribution quantile thresholds;

[0057] The sensor data is divided into training set and validation set using time series slicing strategy, and the prediction feature set is constructed based on the training set and validation set.

[0058] Construct a flight altitude prediction model, train the flight altitude prediction model through the prediction feature set, and predict the flight altitude based on the trained flight altitude prediction model.

[0059] In this embodiment of the present invention, the empirical data used is derived from the flight status captured by the rapid access recorder of a mainstream narrow-body passenger aircraft. The sampling frequency is set to once per second. The raw data matrix includes aerodynamic state indicators (angle of attack), kinematic parameters (ground speed / altitude), and environmental variables (wind speed / direction). To achieve high-precision modeling of the time-series evolution of aircraft altitude, a multidimensional feature engineering framework was constructed. This section constructs an anti-interference feature set based on four types of heterogeneous sensor data: wind direction, wind speed, ground speed, and angle of attack. The specific process is as follows:

[0060] To address data outliers caused by sensor noise and transient disturbances, this study designed a dynamic quantile thresholding (DQT) mechanism based on distribution quantile thresholds. This method effectively suppresses outlier data points caused by sensor failure or sudden meteorological disturbances (such as extreme gust loads) by adaptively adjusting the upper and lower bounds. The formula is as follows:

[0061] (1)

[0062] in, Indicates the height data after cleaning, Represents the original height data, and They represent the 2% and 98% quantiles of the height parameter respectively. This method effectively eliminates the influence of extreme outliers while retaining the main data distribution characteristics.

[0063] Taking into account the dimensional differences of flight parameters and the potential influence of outliers, a robust normalization method is adopted:

[0064] (2)

[0065] in, represents the standardized data, represents the original data points, represents the median of the data set X, Represents the interquartile range of dataset X. This study uses the interquartile range (IQR) as the benchmark for data normalization. Compared with the classic Z-score normalization method, the IQR-based normalization scheme shows better robustness in suppressing data outliers.

[0066] During the data partitioning phase, a time-series slicing strategy was implemented, allocating normalized continuous flight data samples to training and validation sets in a 9:1 ratio. A stratified sampling mechanism was also implemented to maintain covariate balance across the distribution of features across different flight phases. To ensure experimental reproducibility, a baseline random number generator parameter (seed = 42) was set as the foundation for the entire computational process.

[0067] On this basis, a three-stage feature engineering process is used to construct a prediction feature set. The first step is to construct enhanced features based on the core physical characteristics of the aircraft using the following formula:

[0068] (3)

[0069] (4)

[0070] (5)

[0071] in, is the true airspeed of the aircraft, is the wind speed, It's ground speed. is the air density, is the lift coefficient, is the aircraft's angle of attack, is the indicated airspeed.

[0072] Continue to establish time series dynamic characteristics, perform time window statistics on wind speed, ground speed, and airspeed parameters respectively, and smooth the instantaneous fluctuations of wind speed, ground speed and other parameters through the 3-point moving average method to help the model distinguish normal flight status from sudden disturbances and enhance the recognition ability of steady-state flight phase. Use the first-order difference method to quantify the instantaneous change rate of flight parameters, reflect the differential characteristics of the aircraft dynamic state, eliminate the non-stationarity of the time series, and convert the absolute quantity into a change quantity, so that the model can more easily learn the causal relationship between altitude changes and aerodynamic parameter dynamics. Finally, use the 5-point volatility method to measure the oscillation amplitude of flight parameters to characterize the stability of the aerodynamic environment. As a risk warning feature, high volatility suggests that it may enter a stall or turbulent state, guiding the attention mechanism to focus on abnormal periods. The formula is as follows:

[0073] (6)

[0074] (7)

[0075] (8)

[0076] in, is a 3-point moving average. For time point Parameter value, is the first-order difference of parameter x at time t, is the parameter value at time t, is the parameter value at time t-1, is the 5-point volatility at time t, For time Parameter value, is the average value of the parameters in the time window t-4 to t.

[0077] Finally, the wind-to-ground speed ratio is introduced to characterize the aerodynamic environment, quantify the coupling effect between the wind field and the aircraft motion, reveal the energy transfer relationship of the relative airflow, and introduce dimensionless parameters to eliminate the interference of dimensional differences on the model weights and enhance the model's sensitivity to the boundary conditions of the flight envelope, such as the critical wind shear during takeoff and landing. The formula is as follows:

[0078] (9)

[0079] in, is the wind-to-ground speed ratio, is the wind speed, is the ground speed, Represents a very small constant.

[0080] The hybrid feature engineering framework designed in this study, through the coordinated optimization of physical prior knowledge and data-driven principles, enables the transition of aircraft dynamic state deduction from a data-driven black-box model paradigm to a physically constrained white-box model. Systematic evaluation and verification have demonstrated that this architecture surpasses existing baseline models in predictive accuracy, computational efficiency, and physical interpretability. This framework establishes a theoretical framework and provides an engineering implementation solution for the evolution of digital twins for civil aircraft safety monitoring systems.

[0081] The prediction model constructed in this paper is based on a multimodal feature fusion framework, combining a convolutional neural network feature extraction module with an attention mechanism component. The specific structure is shown in Figure 1. The model design fully integrates the time series characteristics of aerodynamic parameters and their dynamic relationship with physical characteristics, and uses a hierarchical feature learning strategy to achieve high-precision flight altitude prediction.

[0082] Multimodal feature processing module

[0083] In view of the heterogeneous characteristics of aviation data, the model adopts a dual-channel processing architecture:

[0084] (1) Time series feature extraction channel: Design a deep convolutional network to process time series features, which is mathematically expressed as:

[0085] (10)

[0086] (11)

[0087] (12)

[0088] in, Represents the first layer convolution output, using 32 groups of 5×1 convolution kernels to extract local fluctuation features. represents a one-dimensional convolution with a kernel size of 5, It is multi-channel time series data; is the sub-layer downsampling feature, Represents a one-dimensional convolution with a kernel size of 3; It is a compressed feature vector generated by adaptive average pooling; the activation function uses GELU; Maxpool realizes nonlinear feature screening by selecting the maximum value in the local window as the output; AdaptiveAvgPool compresses time series features of arbitrary length into vectors of fixed dimension by dynamically calculating the pooling window size; it has better gradient continuity and can enhance the ability to represent time series patterns.

[0089] (2) Physical feature enhancement pathway: A dynamic adjustment mechanism for attention weights is constructed to optimize feature selection based on prior knowledge of flight mechanics, focusing on enhancing the influence coefficients of core aerodynamic variables in the altitude prediction model, thereby improving the transparency of the physical analysis of model decision-making. The calculation process is as follows:

[0090] (13)

[0091] Where, represents the attention weight, σ represents the Sigmoid function, 、 is the learnable weight matrix, is the physical characteristic matrix.

[0092] Feature fusion and regression prediction:

[0093] The features after dual-channel processing are fused through tensor splicing, and the expression is as follows:

[0094] (14)

[0095] in, represents the fused feature vector, Represents the output of the temporal feature extraction channel, and ⊙ represents the Hadamard product operation.

[0096] The fused feature vector is input into the deep regression network composed of GELU activation function and Dropout layer. The specific calculation formula is as follows:

[0097] (15)

[0098] (16)

[0099] (17)

[0100] Where: 、 Corresponding to the primary regression layer output and the secondary regression layer output, respectively, 、 、 The weight matrices corresponding to the primary regression layer, secondary regression layer and output layer are respectively, where the predicted target is a highly predictive value; Dropout is a stochastic regularization technique that drops neuronal connections with a certain probability during training; the GELU function adaptively gates input features through a probability-weighted mechanism. The network architecture employs a layered dimensionality reduction strategy, with the fully connected layer dimension decreasing step by step: 256 → 128 → 1. This effectively mitigates the risk of overfitting while maintaining feature expressiveness. A Dropout mechanism is applied before the final regression layer, achieving regularization by randomly blocking 30% of neuronal connections, significantly improving the model's generalization ability in unknown data domains.

[0101] Experiment and result analysis:

[0102] Experimental configuration: The experimental programming language is Python 3.8. The experiment is conducted on a CPU configured as AMD Ryzen 7 6800HS 3.2GHz, 32GB memory, and 1TB solid-state drive, and the operating system is Windows 11. The Huber loss function is set to δ=50.0, the Adam optimizer, the initial learning rate is 0.01, the L2 regularization coefficient is 0.05, the decay factor is 0.5, and the patience threshold is set to 15 epochs. QAR-1 is the flight data of the landing approach phase collected by QAR for flights operated in 2020. During this phase, the aircraft encountered lateral wind shear disturbances; QAR-2 contains the complete flight parameters of the approach phase recorded under standard atmospheric conditions for a flight in 2021; QAR-3 covers the terminal area flight data captured by the QAR equipment of a flight in 2022. There are lateral airflow disturbances during the approach phase. The experimental dataset information is as follows shown.

[0103] Table 1 Summary of experimental datasets

[0104]

[0105] Prediction Performance Comparison: This paper selects the LSTM architecture as the benchmark comparison model. To ensure consistency of the experimental benchmark, only the topology of the original model is replaced with a recursive architecture composed of LSTM units. The remaining hyperparameter configurations and feature engineering processes remain unchanged. The evaluation system is constructed based on a three-dimensional indicator framework of mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R²). This multi-dimensional performance evaluation framework can systematically analyze the comprehensive performance of the model in terms of error distribution characteristics, stability, and generalization ability. The calculation formulas for the three indicators are:

[0106] (18)

[0107] (19)

[0108] (20)

[0109] in is the true value, is the predicted value, is the mean of the true values.

[0110] Table 2 Figure 2 、 Figure 3 and Figure 4 The performance of the two methods on the test set is shown:

[0111] Table 2 Comparison of prediction performance

[0112] (a) LSTM model

[0113]

[0114] (b) CNN model

[0115]

[0116] Based on Table 2 and Figure 2 、 Figure 3 and Figure 4 Empirical analysis shows that compared to the LSTM model, the proposed method achieves significant improvements in the coefficient of determination dimension, accurately capturing the temporal evolution of flight altitude. Generalization tests across weather conditions and aircraft configurations demonstrate superior predictive robustness: in the crosswind disturbance scenarios represented by the QAR-1 and QAR-3 datasets, the predicted curves are highly consistent with the measured trajectories. Verification on the QAR-2 standard meteorological dataset demonstrates significant reductions in the mean square error (MSE) and mean average error (MAE) by 55.4% and 42.2%, respectively. Joint validation across these three datasets confirms that the multi-source feature fusion architecture effectively improves prediction accuracy and stability through a complementary feature mechanism.

[0117] Ablation experiment:

[0118] This paper verifies the effectiveness of the improved algorithm through ablation experiments. To examine the independent contributions and synergistic mechanisms of each innovative component, the present invention designs a step-by-step ablation experiment framework to systematically deconstruct the marginal effects of modules on prediction performance. The complexity of the flight altitude prediction task stems from the strong coupling effect between multi-source heterogeneous data such as sensor noise and meteorological disturbances and aircraft dynamics. Mainstream methods often rely on end-to-end black-box modeling, which has significant deficiencies in feature engineering interpretability and quantification of architectural component contributions. By progressively integrating flight mechanics feature enhancement modules, attention gating mechanisms, and depthwise separable convolutional architectures, a causal analysis chain of "physical constraints → feature focusing → computational optimization" is constructed to verify the physical interpretability of feature engineering, quantify the focusing ability of the attention mechanism, and demonstrate the effectiveness of network lightweighting.

[0119] Table 3. Comparison of ablation experiment performance

[0120]

[0121] As shown in Table 3 and Figure 5 as well as Figure 6 As shown, the baseline model (using only temporal convolutional channels) achieves a MAE of 155.51 on the QAR-2 dataset. Introducing physical feature engineering reduces this error by 56.1%, demonstrating that explicit modeling of aerodynamic parameters significantly enhances state representation. Further adding the spatiotemporal attention module reduces the MAE to 49.90, validating its effectiveness in suppressing transient noise such as gusts of wind. The complete model ultimately achieves a MAE of 36.05, a 76.8% improvement over the baseline model.

[0122] In addition, Figures 6 and Figure 5 The complete model's prediction error (±82 feet) throughout the final approach phase and at low altitudes below 1,000 feet is significantly lower than the baseline model with physical feature enhancement (±214 feet). This advantage stems from the attention mechanism's dynamic weighting of highly volatile features during the approach phase. Ablation experiments empirically demonstrate that the multi-module collaboration achieves a closed-loop optimization strategy of "physical mechanism guidance, dynamic feature focus, and efficient computational support," providing reliable theoretical support for flight safety monitoring in complex scenarios.

[0123] Conclusion: To address key challenges in flight altitude prediction, including the lack of cross-modal data fusion, insufficient noise immunity, and limited ability to parse time-varying features, this study constructed a hybrid modeling architecture that integrates temporal attention gating with deep convolution operations. Through a collaborative optimization mechanism of multi-scale aerodynamic feature extraction, dynamic allocation of spatiotemporal attention weights, and lightweight network topology, high-precision modeling and dynamic state deduction of the aircraft's altitude temporal evolution were achieved. The main conclusions are as follows:

[0124] 1) Multi-dimensional Feature Collaborative Modeling: Based on flight dynamics principles, an enhanced feature space is constructed that includes parameters such as airspeed, dynamic pressure, and wind-to-ground speed ratio. Hierarchical feature processing (steady-state extraction → transient separation → coupled modeling) is employed to significantly improve the accuracy of dynamic state representation. Experiments show that this strategy reduces the mean absolute error (MAE) by 56.1%.

[0125] 2) Spatiotemporal Attention Optimization: The dual-channel architecture captures short-term fluctuation patterns through convolutional networks and uses an attention mechanism to dynamically assign weights to key parameters. Ablation experiments confirm that the attention module reduces the mean average error (MAE) by an additional 26.9%, and reduces prediction error by 61.7% in low-altitude areas below 1,000 feet compared to traditional convolutional networks.

[0126] 3) Lightweight network design: Using one-dimensional convolution instead of 3D convolution, combined with adaptive learning rate and Huber loss function, it achieves MAE 36.05 feet and R²=0.9966 on the QAR-2 dataset, meeting real-time monitoring needs.

[0127] 4) Physical Interpretability: The distribution of attention weights is highly consistent with aerodynamic theory (e.g., the weights increase significantly during gusty periods), revealing the physical basis for the model's decisions. This method can be integrated into flight safety monitoring systems to provide real-time warnings of altitude deviations, providing reliable technical support for civil aviation safety management and fuel efficiency optimization.

[0128] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A flight altitude prediction method based on a neural network, characterized in that: The method comprises: Collect sensor data from passenger aircraft and suppress outlier data points through a dynamic truncation mechanism based on distribution quantile thresholds; The sensor data is divided into training set and validation set using time series slicing strategy, and the prediction feature set is constructed based on the training set and validation set. Construct a flight altitude prediction model, train the flight altitude prediction model using the prediction feature set, and predict the flight altitude based on the trained flight altitude prediction model; The flight altitude prediction model adopts a multimodal feature fusion framework, including a multimodal feature processing module. The multimodal feature processing module adopts a dual-channel processing architecture, including a time series feature extraction channel: a deep convolutional network is constructed to process time series features. Its mathematical expression is: ； ； ； in, Represents the first layer convolution output, using 32 groups of 5×1 convolution kernels to extract local fluctuation features. represents a one-dimensional convolution with a kernel size of 5, It is multi-channel time series data; is the sub-layer downsampling feature, Represents a one-dimensional convolution with a kernel size of 3; It is a compressed feature vector generated by adaptive average pooling; the activation function uses GELU; Maxpool selects the maximum value in the local window as the output to achieve nonlinear feature screening; AdaptiveAvgPool dynamically calculates the pooling window size to compress time series features of any length into a vector of fixed dimension; Physical feature enhancement pathway: Construct a dynamic adjustment mechanism for attention weights. The calculation process is as follows: ； Where, represents the attention weight, σ represents the Sigmoid function, 、 is the learnable weight matrix, is the physical characteristic matrix.

2. The method for predicting flight altitude based on a neural network according to claim 1, wherein: The sensor data includes at least wind direction, wind speed, ground speed and angle of attack.

3. The method for predicting flight altitude based on neural network according to claim 1, characterized in that: In the step of suppressing outlier data points by a dynamic truncation mechanism based on a distribution quantile threshold, the dynamic truncation mechanism based on a distribution quantile threshold is expressed as: ； in, Indicates the height data after cleaning, Represents the original height data, and They represent the 2% and 98% quantiles of the height parameter, respectively.

4. The method for predicting flight altitude based on neural network according to claim 3, characterized in that: The steps of collecting sensor data from passenger aircraft and suppressing outlier data points through a dynamic truncation mechanism based on distribution quantile thresholds also include normalizing the sensor data, which can be expressed as: ； in, represents the standardized data, represents the original data points, represents the median of the data set X, represents the interquartile range of the data set X.

5. The method for predicting flight altitude based on neural network according to claim 1, characterized in that: The sensor data is divided into training set and validation set by adopting time series slicing partitioning strategy. In the step of obtaining training set and validation set, the data sample ratio of training set and validation set is 9:

1. When dividing validation set and training set, stratified sampling mechanism is adopted to extract sample data.

6. The method for predicting flight altitude based on neural network according to claim 1, characterized in that: In the step of building a prediction feature set based on the training set and the validation set, the prediction feature set is built through three-stage feature engineering. First, the core physical characteristics of the aircraft are enhanced by the following formula: ； ； ； in, is the true airspeed of the aircraft, is the wind speed, It's ground speed. is the air density, is the lift coefficient, is the aircraft's angle of attack, is the indicated airspeed; Time series dynamic characteristics are established. Time window statistics are performed on wind speed, ground speed, and airspeed parameters respectively. The instantaneous fluctuations of wind speed and ground speed are smoothed by the three-point moving average method. The instantaneous change rate of flight parameters is quantified using the first-order difference method, and the absolute value is converted into the change value. The oscillation amplitude of the flight parameters is measured using the five-point volatility method to characterize the stability of the aerodynamic environment as a risk warning feature. The formula is expressed as follows: ； ； ； in, is a 3-point moving average. For time point Parameter value, is the first-order difference of parameter x at time t, is the parameter value at time t, is the parameter value at time t-1, is the 5-point volatility at time t, For time Parameter value, is the parameter average value from time window t-4 to t; The wind-to-ground speed ratio is introduced to characterize the aerodynamic environment, and the formula is expressed as: ； in, is the wind-to-ground speed ratio, is the wind speed, is the ground speed, Represents a very small constant.

7. The method for predicting flight altitude based on neural network according to claim 6, characterized in that: The features after dual-channel processing are fused through tensor splicing, and the expression is: ； in, represents the fused feature vector, Represents the output of the temporal feature extraction channel, and ⊙ represents the Hadamard product operation.

8. The method for predicting flight altitude based on neural network according to claim 7, characterized in that: The fused feature vector is input into the deep regression network composed of GELU activation function and Dropout layer. The specific calculation formula is as follows: ； ； ； Where: 、 Corresponding to the primary regression layer output and the secondary regression layer output, respectively, 、 、 The weight matrices corresponding to the primary regression layer, secondary regression layer and output layer are respectively, where the predicted target is a highly predicted value; Dropout is a random regularization technique that discards neuron connections with a certain probability during training; the GELU function adaptively gates input features through a probability weighting mechanism.

9. The method for predicting flight altitude based on neural network according to claim 1, characterized in that: When collecting sensor data from a passenger aircraft, the sampling frequency is once per second.

Images