Landing safety event prediction method and system based on QAR data and multi-task learning

Abstract

The invention relates to a landing safety event prediction method and system based on QAR data and multi-task learning, and belongs to the technical field of aviation flight safety. The method specifically comprises the following steps: S1, processing multi-parameter QAR data through a multi-scale sharing encoder; s2, performing task specific parameter weight distribution on the shared unified feature representation through a parameter selector; s3, performing time dependence modeling on the task related features through a time sequence decoder; and S4, utilizing a multi-task loss function to jointly optimize the prediction tasks of the plurality of security events. The technical scheme of the invention can be used for simultaneously predicting various safety events in a flight landing stage, such as a tail wiping event and a hard landing event, and a flight safety guarantee tool is provided for an airline company; through deep mining of potential association between flight data and different safety events, pilot training and operation processes are optimized, and flight safety standards are improved.

Description

Technical Field

[0001] The present invention belongs to the field of aviation flight safety technology and relates to a landing safety event prediction method and system based on QAR data and multi-task learning. Background Art

[0002] Aviation safety has always been a core issue in the civil aviation industry. According to the 2024 Annual Report of the International Air Transport Association (IATA), global international passenger traffic is expected to grow by 13.6% year-over-year compared to 2023, while capacity is projected to increase by 12.8%. This significant growth reflects the continued recovery and expansion of the global aviation market, while also placing higher demands on aviation safety. Aviation safety statistics show that, although the landing phase accounts for only a small portion of total flight time, it has a higher accident rate than other phases. During the landing phase, major safety events include hard landings, tail strikes, and runway overruns. These events can not only cause serious damage to the aircraft structure but also pose a significant threat to passenger safety. Currently, the Quick Access Recorder (QAR), an airborne flight data recording device, has been widely used in the civil aviation industry. It continuously records thousands of flight parameters at varying sampling rates, providing critical data support for aviation safety research. With the assistance of QAR data, this paper utilizes a deep multi-task learning approach to jointly predict multiple flight safety events (tail strikes and hard landings) and to explore and explain the potential correlations between these events. In addition, the proposed model is interpretable and can assist in discovering parameters that are highly correlated with safety events, thereby helping pilots improve flight safety.

[0003] In recent years, scholars have conducted extensive research on flight safety, which can be roughly divided into two categories: flight safety research based on expert knowledge analysis and data-driven flight safety incident research.

[0004] Flight safety research based on expert knowledge analysis primarily relies on flight experts' profound insights into civil aviation operations. It explores the risk factors of unsafe incidents from the perspectives of risk factors, crew psychology, and flight management, and further proposes preventive measures. Chen Zhihua et al. analyzed hard landing incidents from the perspective of physical principles and further found that crew psychological factors (fear of hard landings and pursuit of standard landing operation modes) are closely related to hard landings, tail strikes, and runway overruns. They also proposed recommended factors and specific measures for preventing hard landings. Zhou Jin analyzed the patterns of hard landing incidents in recent years, derived the likelihood and impact of each risk factor based on expert scores, and evaluated and ranked the risk factors using the entropy-weighted TOPSIS method. The authors identified the main causes of hard landing incidents and proposed operational and psychological recommendations to address these causes.

[0005] Data-driven research on flight safety events is primarily based on machine learning algorithms. Li Xu et al. used curve clustering to identify hard landing types. They first summarized the causes of hard landing events through group and individual comparative analysis, then established a secondary classification tree. Based on this, they extracted curve features for hard landing pattern identification. Wang Lei et al. expanded the hard landing risk assessment model into a landing risk assessment model, simultaneously considering three flight events: tail strike, hard landing, and runway overrun. Chen et al. used quantile regression to explore the influencing parameters of maximum vertical acceleration during landing. Furthermore, support vector machines and random forests have been used for hard landing event prediction and early warning. Kang Zongwei et al. used a multi-layer spatiotemporal encoder and TG-Attention to predict aircraft landing speed, providing direct warning of runway overrun risks. Tong Chao et al. used an LSTM (long short-term memory) model to predict the vertical load and landing speed of an aircraft during landing, thereby providing early warning of hard landing and runway overrun events. Chen Hongnian et al. also used an LSTM model to predict the pitch angle of an aircraft during landing, providing early warning of tail strike events.

[0006] By jointly optimizing multiple related tasks, multi-task learning models can effectively capture shared features across tasks as well as task-specific features, thereby enhancing the model's generalization ability and improving the predictive performance of each task. Based on the parameter sharing strategy, existing multi-task learning research can be broadly categorized into hard parameter sharing and soft parameter sharing methods. Caruana et al. proposed a hard parameter sharing architecture that shares parameters in lower layers of the model to capture commonalities between tasks, while simultaneously introducing independent output layers at the top of each task to learn task-specific features. Our approach follows a hard parameter sharing architecture, which not only reduces the risk of overfitting but also improves data utilization efficiency through shared representations, performing well when tasks are strongly related. However, it can lead to optimization conflicts when tasks differ significantly. To address this optimization conflict, Misra et al. proposed the Cross-Stitch Network, which explicitly learns shared and task-specific features by introducing learnable "cross-stitch" connections, thereby achieving more flexible information sharing in multi-task learning. Ruder et al. introduced the Sluice Network, which utilizes a layer-by-layer sharing mechanism and gating units to dynamically learn shared and task-specific information, thereby achieving more flexible information sharing between loosely related tasks. Ma et al. proposed MMoE (Multi-gate Mixture-of-Experts) based on MoE (mixture-of-experts). MMoE uses a task-specific gating network to dynamically adjust the weight assignment of each task to the shared expert module, enabling the model to adaptively learn shared features and unique task-specific features across tasks, thereby showing significant performance improvements in multiple tasks. Tang et al. proposed PLE (Progressive Layered Extraction), which explicitly separates experts to reduce the interference between shared knowledge and task-specific knowledge, and uses progressive layered extraction to dynamically adjust the degree of information sharing across tasks. However, multi-task learning models have not yet been applied in previous flight safety research.

[0007] In recent years, with the rapid development of the Internet of Things (IoT), big data, and intelligent technologies, a large number of flight safety research projects based on QAR data have emerged, including data-driven flight safety analysis and aviation accident prediction. Related research based on flight data includes methods based on anomaly detection, traditional machine learning, and deep learning. However, most existing aviation safety incident research, when applied to QAR data, suffers from the following common problems:

[0008] Insufficient prediction accuracy: The sampling frequency of QAR data parameters is not consistent. Many existing methods achieve time series alignment of parameters by taking the average value over a fixed time window, which will undoubtedly lose a lot of time series information and lead to low prediction accuracy.

[0009] Limited consideration of multiple safety events: Existing research has primarily focused on single safety events, ignoring the fact that, in practice, safety risks in different phases of flight are often diverse and interrelated. This narrow focus limits the model's ability to capture the complexity of real-world risk scenarios, reducing the generalizability and practical applicability of the approach. For example, during a low-speed landing, a series of interrelated hazards arise when the aircraft descends below the standard glide path. To avoid a hard landing caused by a stall, the pilot raises the nose, resulting in an excessive pitch angle and increasing the risk of a tail strike. If the pilot increases the airspeed to avoid a tail strike and a hard landing, it may result in a long level-off and may require a go-around to prevent running off the runway.

[0010] Incomplete interpretability: Existing research has primarily focused on improving the prediction performance of flight safety events, while interpretability has been largely overlooked. Recently, Li Xu et al. and Chen Haodong et al. explored the interpretability of hard landing events using class activation mapping and attention mechanisms. However, these approaches only consider explaining a single event, while the correlation and interpretation of multiple flight safety events remain unexplored. Summary of the Invention

[0011] In view of this, the purpose of the present invention is to provide a landing safety event prediction method and system based on QAR data and multi-task learning, which is used to simultaneously predict multiple safety events during the flight landing phase, such as tail strikes and hard landings, to provide airlines with a flight safety assurance tool; by deeply mining the potential correlation between flight data and different safety events, it optimizes pilot training and operating procedures and improves flight safety standards.

[0012] In order to achieve the above object, the present invention provides the following technical solutions:

[0013] A landing safety event prediction method based on QAR data and multi-task learning, the method specifically comprising the following steps:

[0014] S1. Processing multi-parameter QAR data through a multi-scale shared encoder. The multi-scale shared encoder uses convolutional layers with adaptive kernel sizes to independently encode QAR parameters of different sampling frequencies and generate a unified feature representation shared across tasks.

[0015] S2. Assigning task-specific parameter weights to the shared unified feature representation through a parameter selector, wherein the parameter selector generates parameter importance weights based on global average pooling and a learnable convolutional layer, and generates task-related features through weighted selection;

[0016] S3. Modeling the temporal dependency of the task-related features through a temporal decoder. The temporal decoder introduces a fine-grained attention mechanism, treats each parameter at each time step as an independent token input, and combines a learnable global vector to capture the complex temporal relationship across parameters to generate the final prediction result.

[0017] S4. Utilize a multi-task loss function to jointly optimize the prediction tasks of multiple safety events. The loss function automatically adjusts the weight of each task based on task uncertainty and outputs a joint prediction result for tail strike events and hard landing events.

[0018] Furthermore, in step S1, the multi-scale shared encoder is implemented as follows:

[0019] S11. For each QAR parameter Perform an independent one-dimensional convolution operation, and the convolution kernel size and step size are set to the sampling frequency f of the parameter i , the number of output channels is o_c, and high-dimensional encoding features are obtained

[0020] S12. Stack the encoding features of all parameters along the parameter dimension to form a shared feature space

[0021] Furthermore, in step S2, the parameter selector is implemented as follows:

[0022] S21, for shared features x enc Perform global average pooling to obtain the global feature vector

[0023] S22, generate parameter selection weight vector through convolution layer and Sigmoid function And based on the weight, the shared features are weighted and selected to generate task-related features

[0024] Furthermore, in step S3, the timing decoder is implemented as follows:

[0025] S31, task-related features x sel Convert to The Token sequence is concatenated with the learnable global vector and then input into the attention mechanism module;

[0026] S32. Aggregate global information through the multi-head self-attention mechanism and generate the final prediction result based on the updated global vector.

[0027] Furthermore, in step S4, the multi-task loss function is defined as:

[0028]

[0029] Among them, σ p ,σ v represent the uncertainty parameters of the tail strike prediction task and the hard landing prediction task, y p represents the truth value of PITCH, represents the predicted value of PITCH; where y v and represent the true value and predicted value of IVV respectively.

[0030] Furthermore, the method further includes an interpretability analysis step, including: generating a parameter importance heat map based on the weights output by the parameter selector; and analyzing the correlation and risk causes of different security events by visualizing the time series changes of key parameters.

[0031] Furthermore, the QAR data includes at least 17 key parameters including vertical acceleration (VRTG), vertical velocity (IVV), pitch angle (PITCH), roll angle (ROLL), and radio altitude (RADIO_LH / RADIO_RH).

[0032] Furthermore, the prediction results correspond to the flight safety event risk levels at three key moments: 10 feet, 30 feet, and 50 feet before the aircraft lands.

[0033] The present invention also provides a landing safety event prediction system based on QAR data and multi-task learning.

[0034] The beneficial effects of the present invention are:

[0035] 1) To address the frequency-inhomogeneous parameters, a multi-scale shared encoder is proposed, which uses convolutional layers with adaptive kernel sizes to generate unified representations across tasks.

[0036] 2) A gating network-based parameter selector learns task-specific parameter importance, enabling interpretable predictions by identifying key operational features.

[0037] 3) The temporal decoder with a fine-grained attention mechanism captures the complex temporal dependencies between parameters.

[0038] 4) This invention uses a deep multi-task learning model to simultaneously predict multiple flight safety events, providing new application scenarios and research challenges in the field of flight safety.

[0039] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0041] Figure 1 It is the overall architecture of the MultiSafe model;

[0042] Figure 2 It is a multi-scale shared encoder (Multi-scale Shared Encoder);

[0043] Figure 3 is a parameter selector;

[0044] Figure 4 This is the multi-task learning version of the comparison model;

[0045] Figure 5 Heat map of mission parameter importance for hard landing prediction;

[0046] Figure 6 Heat map of the importance of parameters for the tail strike prediction task;

[0047] Figure 7 for flight parameter changes that have high contributions to different prediction tasks;

[0048] Figure 8 This is a visualization diagram of some parameters. DETAILED DESCRIPTION

[0049] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0050] In this embodiment, the specific architecture of the MultiSafe model is as follows: Figure 1As shown in the figure, MultiSafe consists of three components from bottom to top: (1) Multi-scale Shared Encoder: used to process parameters with different sampling frequencies in QAR data to avoid information loss and extract shared features across tasks to enhance the generalization ability of the model; (2) Parameter Selector: designed to obtain task-specific parameter information to meet the needs of different tasks, resolve optimization conflicts, and provide explanation support for flight safety events from the perspective of parameter importance; (3) Temporal Decoder: used to explicitly model the complex temporal relationship between parameters in QAR data and generate prediction results.

[0051] Multi-scale shared encoder:

[0052] The structure of the multi-scale shared encoder is as follows Figure 2 As shown in Figure 2, its core design purpose is to solve the timing misalignment problem caused by different sampling frequencies of multi-parameter QAR data. Specifically, the input data Represents the i-th parameter of QAR data, where f i represents the sampling frequency of the parameter, and t represents the sampling time. Here, i = 1, 2, ..., n, where n is the total number of QAR parameters. Due to the difference in sampling frequency, achieving temporal alignment across parameters is challenging. Previous studies often took the average value over a one-second interval to achieve temporal alignment. However, this approach inevitably leads to the loss of temporal information, which reduces the model's predictive performance.

[0053] In order to make full use of the information in the data, this paper proposes a multi-scale shared feature extraction framework based on Convolutional Neural Networks. This framework independently encodes each parameter into a high-dimensional representation to maintain the integrity of temporal information. Technically speaking, for the parameter sequence x i , compresses its time length to t through one-dimensional convolution operation, while increasing its dimension to maintain the integrity of the information. The kernel size and stride of the convolution are both set to f i , the number of output channels is o_c. After the convolution operation, we get x i High-dimensional representation of Afterwards, we will Stack along the parameter dimension to construct shared encoding features As a unified feature space, shared encoding can not only capture the multi-scale temporal information of all parameters, but also realize cross-mask feature sharing in multi-task learning, thereby enhancing the generalization and prediction capabilities of the model. The calculation steps of the multi-scale shared encoder are defined as follows:

[0054]

[0055] Parameter selector:

[0056] QAR data is inherently multi-parametric, and different safety events rely on different parameters for accurate prediction. For example, predicting a tail strike typically relies more on aircraft attitude parameters such as pitch and roll, while predicting a hard landing relies more on landing impact metrics such as vertical acceleration (VRTG) and vertical velocity (IVV). This variability highlights the need for customized parameter combinations when modeling different flight safety events. A multi-scale shared encoder uses multi-scale convolution operations to process multivariate QAR data, effectively handles inconsistent sampling frequencies, and extracts shared features across tasks. However, it lacks the ability to assign parameter importance based on the specific requirements of each safety event. A multi-scale shared encoder uses multi-scale convolution operations to process multivariate QAR data, effectively handles inconsistent sampling frequencies, and extracts shared features across tasks. However, it lacks the ability to assign parameter importance based on the specific requirements of each safety event. Therefore, it cannot fully capture task-specific features. To address this issue, this paper proposes a parameter selector, a new module designed to identify and learn task-relevant parameter information. This helps reduce optimization conflicts and improve prediction accuracy. By quantifying the importance of each parameter for each task, the parameter selector also provides a transparent and interpretable view of how different variables affect the results. Figure 3 Specifically, by applying global average pooling (GAP), we transform the high-dimensional shared features x enc Mapping to the global eigenvector Global average pooling averages the shared features along the time dimension t and the feature dimension o_c. Its calculation method is defined as:

[0057]

[0058] To further capture the overall distribution of global features, a convolutional layer with kernel size and stride 1 is introduced, where both the input and output have n channels, producing a global representation The definition is as follows:

[0059] v glb =Conv(v gap ) (4)

[0060] Then, the global representation v is transformed into glb Converted to a probability distribution, resulting in a parameter selection weight vector v selEach element of is normalized to a probability value in the range of [0, 1], which quantifies the contribution of each parameter to a specific task, thereby achieving adaptive learning of parameter weights. sel , for the shared feature x enc Perform weighted selection to generate filtered feature representations specific to a single task The above method is calculated as follows:

[0061] v gap =Sigmoid(v glb ) (5)

[0062]

[0063] Timing decoder:

[0064] In QAR data, the correlation between parameters is not only reflected in their interactions at the same moment, but also manifests as cross-time correlation, such as lag and accumulation. The traditional attention mechanism embeds all parameters of a single time step into basic tokens to model the temporal correlation between different time steps. However, this approach not only eliminates the correlation between multiple parameters, but also leads to limited information within the token of a single time step due to the time-lagged nature of the QAR variables, which may hinder the modeling of global temporal correlation. Therefore, without changing the computational architecture of the attention mechanism, the present invention introduces a new input representation to explicitly learn the complex dependencies between parameters at different time steps. Specifically, the features in the parameter QAR are embedded as Convert to Each parameter at each time step can be regarded as a basic Token input, thereby capturing the dependency between any parameter and any time in the QAR data. In addition, a learnable global vector is introduced. And compare it with x along the Token dimension sel Connect to get Then input it into the Attention module. The calculation steps are as follows:

[0065] g=Concat(vec,x sel ) (7)

[0066] Q,K,V=gw q ,gw k ,gw v (8)

[0067]

[0068] in where d q =d k =dv Through linear transformation, the present invention converts Mapped to queries respectively key Sum

[0069] During the computation, the global vector vec interacts with other tokens, effectively aggregating global information from the entire input. Given that vec has no semantic similarity with other tokens, it avoids favoring any particular token. Furthermore, since vec is learnable, it can be optimized during training to adapt to specific tasks, further distinguishing different tasks in multi-task learning. Finally, the present invention only uses vec′, which aggregates global features, to generate predictions through a linear layer, as shown below:

[0070]

[0071] Loss function:

[0072] In multi-task learning, the performance of the model depends largely on the reasonable distribution of loss function weights for each task. However, in the context of multi-task learning for flight safety event prediction, how to effectively distribute the loss weights of different tasks has not been fully studied. Traditional methods often rely on manual hyperparameter tuning to pursue optimal performance, but this process is not only labor-intensive but may also lead to suboptimal solutions. The present invention adopts the following loss function:

[0073]

[0074] Among them, σ p ,σ v represent the uncertainty parameters of the tail strike prediction task and the hard landing prediction task, y p represents the truth value of PITCH, Denotes the predicted value of PITCH. Similarly, for the hard landing prediction task, we where y v and represent the true value and predicted value of IVV respectively.

[0075] In this embodiment, the experimental data and results are as follows:

[0076] 1) QAR data

[0077] In this embodiment, the proposed MultiSafe model is evaluated mainly through two types of experiments. Among them, the first type of experiment is a comparative experiment, which is used to demonstrate the performance improvement effect of MultiSafe over the existing flight safety event prediction methods. The second type of experiment is an ablation experiment, which is divided into module ablation experiments and single-task ablation experiments. The module ablation experiment is used to verify the effectiveness of the multi-scale shared encoder (Multi-scale Shared Encoder), parameter selector (Parameter Selector) and temporal decoder (Temporal Decoder) in MultiSafe. The single-task ablation experiment is used to explore the impact of multi-task learning on a single task.

[0078] The dataset used in this embodiment contains 37,518 QAR samples of A320 flights landing at domestic airports in China. Each flight sample records multiple parameters throughout the entire flight phase. This embodiment selects 17 key parameters for prediction, including VRTG, IVV, PITCH, ROLL, etc. Table 1 provides a detailed description of the selected parameters and their sampling frequency. To ensure the effectiveness of model training and evaluation, the dataset is divided into training set, validation set, and test set in a ratio of 7:1.5:1.5. For the prediction task, three key moments are considered, corresponding to the time when the aircraft is at an altitude of 10 feet, 30 feet, and 50 feet before landing. Then, the data 20 seconds before the critical moment is used as input to predict the flight safety event at the landing moment. The three experimental schemes are represented as scenarios I, II, and III, respectively.

[0079] Table 1 Description of selected QAR parameters

[0080]

[0081]

[0082] 2) Comparative experiment

[0083] In order to evaluate the prediction performance of MultiSafe, this example selected the following six comparison models for comparative experiments, including Random Forest (RF), LSTM, MLSTM-FCN, DLinear, Transformer, and SDTAN. Figure 4 As shown in Figure 2, a multi-task learning version (MT-) of the above model is constructed by adding an additional prediction head to the output layer of the single-task model and trained using Equation (11) as the loss function.

[0084] For each model, hyperparameters were optimized using cross-validation, and model performance was comprehensively evaluated using mean squared error (MSE) and mean absolute error (MAE). Comparative experimental results (shown in Table 2, with the best results highlighted in bold) demonstrate that MultiSafe demonstrates significant advantages in the joint prediction task of tail strike and hard landing events, achieving the best prediction accuracy across various experimental scenarios. MT-RF, as a traditional machine learning model, exhibits significantly lower prediction accuracy compared to other deep learning models, suggesting that its ability to capture useful features from complex multivariate time series is limited due to its low model complexity. While MT-LSTM benefits from its inherent structural advantages in processing time series, its inability to effectively handle multi-scale data leads to information loss and lower prediction accuracy. MT-MLSTM-FCN introduces FCN on top of LSTM to extract local features, better handling the complex features of multivariate time series and thus achieving higher prediction performance. Due to significant differences in the trend strength of various parameters in the QAR data, and the increasingly unstable trends with increasing distance from the ground, MT-DLinear performed poorly, even falling behind MT-RF in hard landing prediction accuracy in scenario III. MT-Transformer relies on a self-attention mechanism to achieve global temporal modeling, but it struggles to learn correlations between multiple parameters, resulting in poor multi-task prediction performance. MT-SDTAN outperforms other comparison models due to its ability to effectively handle multi-scale data and process both global and local features. Furthermore, it is clear that the prediction performance of all models gradually decreases with increasing distance from the ground, especially when the distance changes from 10 feet to 30 feet, where the performance drop is most significant. This suggests that the closer the distance to the ground, the more critical discriminative information is contained in the data.

[0085] Table 2 Performance comparison of different models

[0086]

[0087] 3) Ablation experiment

[0088] To comprehensively evaluate the effectiveness of the model, this example designs two sets of ablation experiments: in the first set, each module is deleted in turn to verify their contribution to the model performance, while in the second set, all modules are retained but the focus is on single-task prediction to explore the impact of multi-task learning on a single task.

[0089] Module ablation experiments: In the module ablation experiments, the multi-scale shared encoder, parameter encoder, and temporal decoder modules are removed separately while maintaining the model integrity through alternative methods. Specifically, after removing the multi-scale shared encoder module, the parameters are aligned by averaging the values within a one-second interval and processed independently using convolutional layers with a kernel size and stride of 1, an input channel of 1, and an output channel of o_c. They are then stacked along the parameter dimension and input into the next module to produce the ablation model MultiSafe enc- After removing the parameter selector module, we get MultiSafe sel- Finally, by removing the temporal decoder module, the output of the parameter transformer is transformed using a convolution operation with a kernel size and stride of 1, input channels of n×t, and output channels of 1, followed by a linear transformation to generate predictions, denoted as MultiSafe dec- .

[0090] Table 3. Module ablation experiment results

[0091]

[0092]

[0093] Table 3 lists the results of the module ablation experiment. The results show that in all experimental scenarios, MultiSafe enc- The MSE and MAE of the two tasks are significantly higher than those of the complete model. This shows that the multi-scale shared encoder solves the misalignment in the time dimension caused by the different sampling frequencies in the multi-parameter QAR data through multi-scale convolution operations. It avoids the loss of temporal information caused by traditional alignment methods, thereby significantly improving the prediction performance of the model. sel- , all indicators except MAE V in scenario I show different degrees of increment, indicating that the parameter selector realizes task-specific parameter selection and learning based on different safety event characteristics, thereby alleviating the optimization conflict between tasks and improving the prediction accuracy. dec- It outperforms MultiSafe only in the hard landing task under Scenario II. In all other cases, it underperforms MultiSafe. This demonstrates that the temporal decoder plays a crucial role in modeling complex temporal relationships. Overall, all three modules contribute to the predictive performance of the model's multi-scale convolutional operations.

[0094] Single-task ablation experiment: MultiSafe ivv and MultiSafe pit They represent the hard landing prediction task and the tail strike prediction task respectively. The loss function for each single task is expressed as:

[0095]

[0096] As shown in Table 4, MultiSafe demonstrates superior prediction performance compared to the single-task model in all experimental scenarios. This demonstrates that MultiSafe effectively exploits the correlations between different flight safety events through a shared information mechanism, thereby improving the overall prediction performance of the joint prediction task.

[0097] Table 4 Single-task ablation experiment results

[0098]

[0099] Interpretability Analysis:

[0100] To analyze and explain flight safety events, this example uses a high-risk example for a case study. The true IVV is -11.4 ft / s and the true PITCH is 6.7°, indicating a possible hard landing and tail strike. This example is fed into the trained MultiSafe model for evaluation. As shown in earlier experiments, input data from lower altitudes carries more critical information for accurate predictions. Therefore, interpretability analysis is performed under Scenario I. In this case, the model predicts an IVV of -10.7 ft / s and a PITCH of 6.3°, which are very close to the true values.

[0101] Figure 5 and Figure 6 The task-specific parameter weights for hard landing and tail strike prediction are given separately. For the hard landing prediction task, the top five parameters with the highest weights are vertical velocity (IVV), indicated airspeed (IAS), pitch angle (PITCH), right radio altitude (RADIO_RH) and standard corrected altitude (ALT_STD). These parameters have significantly higher weights compared to other parameters, indicating their major impact on the prediction results. For tail strike prediction, the most influential parameters are PITCH, ALT_STD, pitch control (PITCH CMD), vertical acceleration (VRTG) and left radio altitude (RADIO_LH). An interesting observation is that the contributions of right radio altitude and left radio altitude parameters are opposite for the two prediction tasks. This may be due to their high correlation, as Figure 7 As shown, they contain redundant information. Therefore, each task tends to prioritize only one of these parameters to avoid redundancy and focus on the parameters that provide more relevant information for its specific prediction target.

[0102] In the hard landing prediction task, IVV receives the highest weight because it is the target variable being predicted. According to the aircraft lift calculation formula:

[0103]

[0104] Where ρ is the air density, C s is the lift coefficient, v is the aircraft airspeed, and the aircraft lift, Y, is proportional to the square of the airspeed. Since Y is a determinant of IVV, it is reasonable that IAS has the second-largest parameter weight when predicting IVV. Similarly, for the tail strike prediction task, parameters directly related to the aircraft's attitude and position, including PITCH, PITCH_CMD, ALT_STD, and RADIO_LH, are given higher weights. These parameters are closely related to the aircraft's orientation and control behavior during landing, making them crucial for accurate tail strike prediction.

[0105] from Figure 5 and Figure 6 It can be seen that certain parameters contribute significantly to both prediction tasks, indicating that there is an intrinsic relationship between tail strike and hard landing events. Figure 7 Further supporting this correlation, Figure 7 The variations of shared parameters in the case samples are illustrated. It is noteworthy that IVV and PITCH show a clear inverse correlation. This suggests that during the landing phase, excessive adjustments to the pitch attitude (which affects the lift coefficient) to avoid the risk of a tail strike may lead to a hard landing or even an airborne stall event. Conversely, prioritizing a soft landing while ignoring changes in the aircraft attitude will greatly increase the risk of a tail strike. In addition, since RADIO_RH, RADIO_LH, and ALT STD contain a lot of information about the vertical position of the aircraft, they also play an important role in predicting abnormal landing patterns.

[0106] To further investigate the dynamics of tail strike and hard landing events, this example compares key parameters of the case samples and normal landing samples in the test set. Figure 8 As shown, the red curve represents the temporal variation of a specific parameter in the case sample, while the blue shaded area represents the typical range observed in normal landings. Figure 8 (b) and Figure 8 (c) shows that the aircraft maintained a relatively normal altitude for the first 10 seconds. However, starting at t = 10s, its altitude deviated from the normal range. While most normal flights descended to 50 feet (usually the runway entrance point) between t = 14s and t = 16s, the case flight did not reach this altitude until about t = 18s. It is worth noting that all flights, including this case, descended to 10 feet at almost the same time, indicating that the case flight executed a rapid descent after entering the runway. Figure 8As shown in (e), the IVV began to drop below the normal range at t = 6s and reached its minimum at t = 10s. It then gradually increased and returned to the normal range at t = 14s. However, the long period of low IVV caused a deviation in the altitude trajectory. To correct this, the aircraft increased its IVV sharply, eventually exceeding the normal value and reaching its peak at t = 18s. At this point, the descent rate was too great and there was not enough time to slow down before landing. When the aircraft descended to 10 feet, its IVV remained at a high level of about 12 ft / s and was still 11.4 ft / s when it landed. Figure 8 (f) and Figure 8 (g) It further showed that in the final seconds, the aircraft performed a strong flare maneuver, rapidly increasing the pitch angle in an attempt to reduce the IVV. This action caused the pitch to exceed the normal level at 10 feet and exceeded the safety threshold for touchdown, directly leading to the risk of a tail strike.

[0107] Therefore, during the landing phase, to minimize the risk of a safety event, pilots should ensure that the IVV remains at a reasonable level. This prevents the aircraft from entering an abnormal altitude mode, which could result in insufficient margin for corrective action. If there is sufficient altitude to adjust the aircraft, pilots should also strive to avoid abrupt control inputs, such as sudden changes in pitch or roll commands. A more effective strategy is to adjust the aircraft's attitude gradually and incrementally. This step-by-step approach allows for more precise control of pitch and IVV, helps prevent sudden attitude changes, and reduces the likelihood of a tail strike or hard landing.

[0108] The technical solution of this invention is used to simultaneously predict multiple safety events during the landing phase of a flight, such as tail strikes and hard landings, providing airlines with a flight safety assurance tool. By deeply exploring the potential correlations between flight data and different safety events, pilot training and operational processes can be optimized, thereby improving flight safety standards. In summary, MultiSafe provides an effective solution for flight safety event warnings, playing a vital role in improving aviation safety.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A landing safety event prediction method based on QAR data and multi-task learning, characterized by: The method specifically comprises the following steps: S1. Processing multi-parameter QAR data through a multi-scale shared encoder. The multi-scale shared encoder uses convolutional layers with adaptive kernel sizes to independently encode QAR parameters of different sampling frequencies and generate a unified feature representation shared across tasks. S2. Assigning task-specific parameter weights to the shared unified feature representation through a parameter selector, wherein the parameter selector generates parameter importance weights based on global average pooling and a learnable convolutional layer, and generates task-related features through weighted selection; S3. Modeling the temporal dependency of the task-related features through a temporal decoder. The temporal decoder introduces a fine-grained attention mechanism, treats each parameter at each time step as an independent token input, and combines a learnable global vector to capture the complex temporal relationship across parameters to generate the final prediction result. S4. Utilize a multi-task loss function to jointly optimize the prediction tasks of multiple safety events. The loss function automatically adjusts the weight of each task based on task uncertainty and outputs a joint prediction result for tail strike events and hard landing events.

2. The landing safety event prediction method based on QAR data and multi-task learning according to claim 1, characterized in that: In step S1, the multi-scale shared encoder is implemented as follows: S11. For each QAR parameter Perform an independent one-dimensional convolution operation, and the convolution kernel size and step size are set to the sampling frequency f of the parameter i , the number of output channels is o_c, and high-dimensional encoding features are obtained S12. Stack the encoding features of all parameters along the parameter dimension to form a shared feature space 3. The landing safety event prediction method based on QAR data and multi-task learning according to claim 2, characterized in that: In step S2, the parameter selector is implemented as follows: S21, for shared features x enc Perform global average pooling to obtain the global feature vector S22, generate parameter selection weight vector through convolution layer and Sigmoid function And based on the weight, the shared features are weighted and selected to generate task-related features 4. The landing safety event prediction method based on QAR data and multi-task learning according to claim 3, characterized in that: In step S3, the timing decoder is implemented as follows: S31, task-related features x sel Convert to The Token sequence is concatenated with the learnable global vector and then input into the attention mechanism module; S32. Aggregate global information through the multi-head self-attention mechanism and generate the final prediction result based on the updated global vector.

5. The landing safety event prediction method based on QAR data and multi-task learning according to claim 4, characterized in that: In step S4, the multi-task loss function is defined as: Among them, σ p ,σ v represent the uncertainty parameters of the tail strike prediction task and the hard landing prediction task, y p represents the truth value of PITCH, represents the predicted value of PITCH; where y v and represent the true value and predicted value of IVV respectively.

6. The landing safety event prediction method based on QAR data and multi-task learning according to claim 5, characterized in that: The method further includes an interpretability analysis step, including: generating a parameter importance heat map based on the weights output by the parameter selector; and analyzing the correlation and risk causes of different security events by visualizing the time series changes of key parameters.

7. The landing safety event prediction method based on QAR data and multi-task learning according to claim 6, characterized in that: The QAR data includes at least 17 key parameters including vertical acceleration (VRTG), vertical velocity (IVV), pitch angle (PITCH), roll angle (ROLL), and radio altitude (RADIO_LH / RADIO_RH).

8. The landing safety event prediction method based on QAR data and multi-task learning according to claim 7, characterized in that: The prediction results correspond to the flight safety event risk levels at three critical moments: 10 feet, 30 feet, and 50 feet before the aircraft lands.

9. A landing safety event prediction system based on QAR data and multi-task learning, characterized by: The system adopts the method according to any one of claims 1 to 8.

Images