Runway risk prediction model based on improved SCSO algorithm

Abstract

The invention discloses an improved SCSO algorithm-based runway risk prediction model, the runway risk prediction model is mainly composed of a runway risk quantification model and an XGBoost-LightGBM-LSTM fusion prediction model, monthly risk data of the runway is obtained by using the runway risk quantification model, and the monthly risk data of the runway is obtained by using the XGBoost-LightGBM-LSTM fusion prediction model. The fusion prediction model is optimized through an improved sand dune cat optimization algorithm, and compared with the prior art, the impact and deviation runway risk prediction model constructed by the invention realizes accurate quantification and active early warning of flight operation risks; a nonlinear adaptive mechanism, dynamic visual simulation, population information exchange and elite perturbation strategy are innovatively introduced, the global search and local development capability of the sand dune cat algorithm is remarkably improved, the problem that a traditional optimization algorithm is prone to falling into local optimum is effectively solved, and prediction precision and robustness are remarkably improved through multi-algorithm fusion.

Description

Technical Field

[0001] This invention relates to the field of aviation technology, and in particular to a runway overrun risk prediction model based on an improved SCSO algorithm. Background Technology

[0002] Runway overrun refers to an incident during takeoff or landing where an aircraft fails to remain within the runway surface, thus overrunning the runway terminus or veering off the side of the runway. According to data from the International Air Transport Association (IATA) and the Flight Safety Foundation, runway overrun has long been one of the major types of civil aviation incidents, with potential consequences including aircraft damage, personal injury, runway closure, and significant economic losses. Traditional safety management relies heavily on qualitative analysis and post-incident investigations. However, with the widespread adoption of technologies such as QAR (Quick Access Recorder), in-depth mining and quantitative analysis of flight data has become possible. Summary of the Invention

[0003] The purpose of this invention is to provide a runway overrun risk prediction model based on an improved SCSO algorithm. By quantifying the risk of runway overrun events, discrete event characteristics are transformed into a unified hazard index, enabling risk classification, comparison, and trend analysis. The risk prediction model is used to predict runway overrun risks, thereby transforming passive response into proactive early warning and improving the scientific and forward-looking nature of safety management.

[0004] To achieve the above objectives, the technical solution adopted by this invention is: a runway deviation risk prediction model based on an improved SCSO algorithm. This runway deviation risk prediction model mainly consists of a runway deviation risk quantification model and an XGBoost-LightGBM-LSTM fusion prediction model. The specific construction method is as follows. Step 1: Construct a runway deviation risk quantification model. This model identifies the core parameters directly related to runway deviation risk from QAR data sources through event analysis, industry reports, and expert experience. It then scores and assigns values ​​to these parameters to quantify them. The total risk index (TRI) is obtained by summing the scores of all identified parameters. Finally, the runway deviation risk dataset is obtained by sequentially aggregating the runway deviation risks for each time period according to the time series. Step 2: Improve the sand cat optimization algorithm mechanism; in the prey search stage, introduce a nonlinear adaptive sensitivity parameter to replace the original linear decreasing mechanism; in the prey attack stage, introduce a dynamic perspective simulation mechanism to simulate the sand cat's visual perception; introduce a population information exchange mechanism in the balance parameter to enhance local exploitation ability; and introduce elite guidance and perturbation strategies to avoid premature convergence. Step 3: Construct a fusion prediction model that includes an LSTM model, a LightGBM model, and an XGBoost model. The fusion prediction model is trained using a runway deviation risk dataset for time series data prediction. Step four: Optimize the fusion prediction model using the improved Sand Cat optimization algorithm; Step 5: Evaluate and test the runway deviation risk prediction model.

[0005] Preferably, in step one, the specific quantification method of the runway deviation risk quantification model is as follows: First, the QAR data is collected and processed: Secondly, different risk index scores are assigned to each parameter according to different states. The specific method is as follows: The status types of key parameters are classified, and safety experts or trained analysts determine the status of each parameter according to the rules of the risk quantification table and select the corresponding risk index score. In cases that are ambiguous or difficult to judge, the principle of conservatism should be followed or the matter should be submitted to the expert group for review. Then, the scores corresponding to the specific states of each parameter involved in the event are added together to obtain the total risk index (TRI) of the event. Finally, to better distinguish and manage runway deviations, a runway deviation risk level threshold is set, dividing it into four levels: red, orange, yellow, and blue.

[0006] As a preferred method, the specific methods for collecting and processing QAR data are as follows. Step a. Event identification: Identify possible runway overrun events through channels such as QAR overrun event reports, crew reports, and ATC reports; Step b. Data extraction: Extract radio altitude, vertical velocity, air-to-ground signal, acceleration, throttle position, brake pressure, thrust reverse status and related parameters from QAR data; Step c. Multi-source information fusion: Integrate meteorological data, NOTAM, runway condition reports, ATC call recordings, and crew cabin voice recordings to accurately assess runway conditions, meteorological conditions, crew decisions, and air traffic control interactions.

[0007] Preferably, the risk level classification for runway deviation is as follows: TRI≥90, extremely high risk, indicating a high probability of an accident occurring. Orange level: 75≤TRI<90, high risk, requires serious attention; Yellow level: 60≤TRI<75, medium risk, used as a safety early warning information for daily risk monitoring and trend analysis; Blue level: TRI < 60, low risk, used as a safety early warning information for daily risk monitoring and trend analysis.

[0008] Preferably, in step two, the improved sandcat optimization algorithm is as follows: 1) Searching for prey During the prey-hunting phase, the sand cat will search for the best prey within its search range and move towards that prey's location, with the location update as shown in the following formula: (1); In the formula, Let c be the initial position of the c-th sand cat during the (t+1)-th iteration. For the sand cat's sensitivity, The optimal individual in generation t; A random number between 0 and 1; Among them, the sand cat's sensitivity The formula is as follows , (2; In the formula, For nonlinear adaptive sensitivity parameters, A random number between 0 and 1; Nonlinear adaptive sensitivity parameters The formula is as follows (3); In the formula, represents the auditory characteristics of the sand cat, and its value is a constant, usually set to 2; T is the maximum number of iterations. This is a non-linear adjustment factor, typically ranging from 1.5 to 2.5; 2) Attacking prey During this phase, each sand cat approaches the hunting position in space along different circumferential directions using a dynamic perspective. The position update formula is shown below: (4); In the formula: The distance is random. For dynamic perspective; The numbers are randomized and randomly perturbed to ensure that the algorithm has a certain degree of randomness, which helps to avoid getting trapped in local optima. Let c be the current position of the c-th sand cat during the t-th iteration; Among them, dynamic perspective The formula is as follows: d (5); In the formula, For maximum sensing distance, This is the distance to the optimal solution at the current distance. d represents a random number that is uniformly distributed in the range [0, 1); 3) Balance parameter R Population information exchange is introduced into the equilibrium parameter R, and the specific formula is as follows: (6); In the formula, This represents the current average position of the population. Weighting for information exchange; like If the sand cat enters the exploration phase, its next location can be a random combination of its prey and its current location, allowing for a global search; if The sand cat will enter the attack and exploitation phase, conducting localized searches; By accelerating algorithm convergence through "elite guidance" and maintaining population diversity through "adaptive perturbation," the algorithm's development and exploration capabilities are effectively balanced, thus avoiding premature convergence.

[0009] Preferably, the elite guidance mechanism involves guiding ordinary individuals in the population towards the current elite individuals (i.e., the current global optimal solution). By aligning the location information of the population, the population can be guided to quickly focus on potentially better regions, thereby improving the convergence speed. The specific method is as follows: The elite guidance item has been incorporated into the position update formula during the prey attack phase. The modified formula is as follows: ; In the formula, The global historical optimal position in the t-th iteration is the elite individual. The elite-leading factor is a constant, typically set to 0.1–0.3. The perturbation strategy involves applying controlled random perturbations to elite individuals to inject diversity into the population and prevent premature convergence caused by excessive clustering of all individuals in a local optimum. The specific method is as follows: When the algorithm detects the globally optimal solution after K consecutive iterations The improvement amount is less than the preset threshold At that time, it was determined that the situation had stalled, and then... Apply a Gaussian perturbation, and the new position after the perturbation. The calculation is as follows: ; In the formula: It is a random number that follows a standard normal distribution, providing the direction and magnitude of the random perturbation; It is the dynamically decaying disturbance intensity coefficient, and its calculation formula is: ,in For initial strength, It represents the range of the search space in each dimension, and is used to scale the perturbation amplitude to adapt it to the problem scale.

[0010] Preferably, in step four, the improved Sand Cat optimization algorithm optimizes the fusion prediction model using the following method: Step A: Obtain the risk dataset for a certain aircraft model for a specific time period based on the runway deviation risk quantification model, and normalize the risk data. Step B: Initialize the improved ISCSO algorithm for the sand cat population X, set the maximum number of iterations T, the nonlinear adjustment factor α, the maximum perception distance D, the information exchange weight γ, and the relevant parameters of the elite guidance strategy, and optimize the XGBoost-LightGBM-LSTM fusion prediction model. In step B, each sand cat performs exploration and development behaviors and updates its location based on the improved ISCSO algorithm, which combines four mechanisms: nonlinear adaptive parameters, dynamic perspective simulation, population information exchange, elite guidance and perturbation. Step D: After each iteration, calculate the RMSE value of the XGBoost-LightGBM-LSTM fusion model corresponding to the parameter combination represented by each sand cat, and update the global optimal solution xbest and the corresponding optimal parameter combination. Step E: Determine if the maximum number of iterations T has been reached: if so, output the optimal parameter combination; otherwise, return to continue iterating. Step F uses the fitness function RMSE as an evaluation metric to guide the ISCSO algorithm in finding the optimal parameter combination in the search space. Step G: After optimization, the optimal parameters obtained in step F are assigned to the XGBoost, LightGBM, and LSTM models respectively to complete the optimization of the runway deviation risk prediction model.

[0011] As a preferred option, the specific method for obtaining risk data in step A is as follows: 1) Retrieve runway departure incident data for the corresponding month for a specific aircraft model, obtain risk data through a runway departure risk quantification model, and calculate the monthly risk value for that aircraft model using the monthly risk value formula, as follows: ; In the formula: i represents the "runaway" event in each month, from 1 to n; 2) To ensure that the diversity of different feature data value ranges and units does not adversely affect model training, the corresponding monthly average risk values ​​are normalized using the following formula: (12); in , It is the number of eigenvalues; , These are the values ​​before and after normalization, respectively. , The first The minimum and maximum values ​​of the feature before normalization.

[0012] Preferably, in step E, the evaluation metric uses the root mean square error of the test set as the fitness value, and the calculation formula is as follows: (11); In the formula, The number of samples; The actual value; These are predicted values.

[0013] As a preferred option, in step five, the optimized XGBoost-LightGBM-LSTM model is tested using test set data, and the RMSE, MAE, and other indicators of the prediction results are calculated to evaluate the model performance and plot the prediction curve and the comparison chart of the true value.

[0014] Compared with the prior art, the advantages of the present invention are as follows: (1) This invention constructs a runway deviation risk prediction model based on an improved SCSO algorithm to optimize the XGBoost-LightGBM-LSTM fusion model, achieving accurate quantification and proactive early warning of flight operation risks. It innovatively introduces nonlinear adaptive mechanisms, dynamic visual simulation, population information exchange, and elite perturbation strategies, which significantly improves the global search and local exploitation capabilities of the Sand Cat algorithm and effectively overcomes the problem that traditional optimization algorithms are prone to getting trapped in local optima.

[0015] (2) This model comprehensively utilizes the advantages of XGBoost, LightGBM, and LSTM under different data structures, and significantly improves prediction accuracy and robustness through multi-algorithm fusion. Compared with single models or traditional scoring methods, this method has stronger interpretability, higher automation, and better generalization ability, and can provide airlines and regulatory agencies with scientific and systematic risk management tools, which has important practical application value for improving flight safety. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is the structural expansion of the LSTM model of this invention. Detailed Implementation

[0017] A runway miss risk prediction model based on an improved SCSO algorithm is proposed. This model mainly consists of a runway miss risk quantification model and an XGBoost-LightGBM-LSTM fusion prediction model. The specific construction method is as follows (see [link]). Figure 1 : Step 1: Construct a runway excursion risk quantification model. This model uses event analysis, industry reports (such as the IATA Runway Excursion Risk Reduction Toolkit), and expert experience to identify core parameters directly related to runway excursion risk from QAR data sources. These parameters cover multiple dimensions, including approach and landing status, aircraft control, external environment, crew performance, and air traffic control intervention. The parameters are scored and quantified, and the total risk index (TRI) is obtained by summing the scores of all identified parameters. The runway excursion risk dataset is then collected sequentially according to the time series for each time period.

[0018] In step one, the specific quantification method for the runway deviation risk quantification model is as follows: First, the specific methods for collecting and processing QAR data are as follows: Step a. Event identification: Identify possible runway overrun events through channels such as QAR overrun event reports, crew reports, and ATC reports; Step b. Data extraction: Extract relevant parameters from QAR data, such as radio altitude, vertical velocity, air-to-ground signal, acceleration, throttle position, brake pressure, thrust reverse status, etc. Step c. Multi-source information fusion: Integrate meteorological data (METAR / SPECI), NOTAM, runway condition reports, ATC call recordings, and crew cabin voice recordings to accurately assess runway conditions, meteorological conditions, crew decisions, and air traffic control interactions.

[0019] Secondly, different risk index scores are assigned to each parameter according to different states. The specific method is as follows: The status types of key parameters should be categorized, and safety experts or trained analysts should determine the status of each parameter according to the rules of the risk quantification table, and select the corresponding hazard index score. For ambiguous or difficult-to-determine situations, a conservative approach should be followed or the matter should be submitted to an expert panel for review.

[0020] The risk quantification table for runway deviation is shown in Table 1.

[0021]

[0022]

[0023]

[0024] The table details the content of each parameter (type), its different states, and their corresponding hazard indices. The data source is primarily QAR records, supplemented by cross-validation of information from cockpit voice recorders, weather intelligence, and ATC call logs.

[0025] The following principles must be followed when setting the scores: Causal correlation: The strength of the correlation between the state of this parameter and the consequences of running off the runway; Severity gradient: The more dangerous the parameter state, the higher the value is assigned, and there is a reasonable gradient between the scores; Data availability: Parameter status should be able to be determined through objective data such as QAR data, cabin voice records, weather reports, and air traffic control records.

[0026] Then, the scores corresponding to the specific states of each parameter involved in the event are added together to obtain the Total Risk Index (TRI) of the event.

[0027] Finally, based on industry safety standards and management requirements, all runway overruns are considered serious incidents in transport aviation. To better differentiate and manage runway overruns, runway overrun risk level thresholds are set, categorizing them into four levels: red, orange, yellow, and blue. Red level: TRI≥90, extremely high risk, indicating a very high probability of an accident occurring; Orange level: 75≤ TR<90, high risk, requires serious attention; Yellow level: 60≤TRI<75, medium risk, used as a safety early warning information for daily risk monitoring and trend analysis; Blue level: TRI < 60, low risk, used as a safety early warning information for daily risk monitoring and trend analysis.

[0028] The traditional Sand Cat Swarm Algorithm (SCSO) consists of two phases: a prey search phase and a prey attack phase. The balance between the two phases is achieved through a balancing parameter mechanism.

[0029] The core idea of ​​the runway deviation risk quantification model is that the severity (hazard index) of a runway deviation event can be obtained by summing the severity scores of a series of key influencing factors (parameters). The construction of the quantification model mainly includes three steps: key parameter identification, parameter score assignment, and hazard index calculation and level classification.

[0030] Application scenarios of the runway deviation risk quantification model: 1. Event severity assessment: Provides objective and quantitative evidence for event investigation and characterization, reducing subjective differences; 2. Safety Trend Analysis: Regularly calculate the risk index of similar events, track the changing trends of risk values ​​for specific parameters, and identify systemic risks; 3. Targeted improvements: Conduct in-depth analysis of high-frequency and high-scoring risk parameters and develop targeted mitigation measures (such as revising SOPs, strengthening training in specific subjects, and improving airport facilities). 4. Pilot training: High-risk event cases and their quantitative analysis results are used in pilot training to improve risk awareness.

[0031] Step 2: Improve the Sand Cat Optimization Algorithm (ISCSO) mechanism; in the prey search phase, introduce a nonlinear adaptive sensitivity parameter to replace the original linear decreasing mechanism; in the prey attack phase, introduce a dynamic perspective simulation mechanism to simulate the Sand Cat's visual perception; introduce a population information exchange mechanism in the balance parameters to enhance local exploitation capabilities; and introduce elite guidance and perturbation strategies to avoid premature convergence.

[0032] The improved Sand Cat Optimization Algorithm (ISCSO) mechanism of this invention is as follows: 1) Searching for prey During the prey-hunting phase, the sand cat will search for the best prey within its search range and move towards that prey's location, with the location update as shown in the following formula: (1); In the formula, Let c be the initial position of the c-th sand cat during the (t+1)-th iteration. For the sand cat's sensitivity, The optimal individual in generation t; A random number between 0 and 1; Among them, the sand cat's sensitivity The formula is as follows , (2; In the formula, For nonlinear adaptive sensitivity parameters, A random number between 0 and 1; Nonlinear adaptive sensitivity parameters The formula is as follows (3); In the formula, represents the auditory characteristics of the sand cat, and its value is a constant, usually set to 2; T is the maximum number of iterations. This is a non-linear adjustment factor, typically ranging from 1.5 to 2.5; In the traditional sand cat swarm algorithm For linear sensitivity parameters, the formula is: The sensitivity parameter decreases linearly, decreasing from 2 to 0 during the iteration process. Compared with the traditional sand cat swarm algorithm, the nonlinear adaptive sensitivity parameter design of this invention makes the transition between exploration and development smoother, enhances the convergence accuracy in the later stages, and the nonlinear decrease is more in line with biological behavior, thus improving the algorithm's adaptability.

[0033] 2) Attacking prey During this phase, each sand cat approaches the hunting position in space along different circumferential directions using a dynamic perspective. The position update formula is shown below: (4); In the formula: The distance is random. For dynamic perspective; The numbers are randomized and randomly perturbed to ensure that the algorithm has a certain degree of randomness, which helps to avoid getting trapped in local optima. Let c be the current position of the c-th sand cat during the t-th iteration; Let c be the current position of the c-th sand cat at the (t+1)-th iteration. Among them, dynamic perspective The formula is as follows: d (5); In the formula, For maximum sensing distance, This is the distance to the optimal solution at the current distance. d represents a random number that is uniformly distributed in the range [0, 1); And traditionally, each sand cat The algorithm uses random angles to approach the hunting position along different circular directions in space. Compared with the traditional sand cat swarm algorithm, this invention utilizes a dynamic perspective. It can better simulate the visual adjustment mechanism of the sand cat, becoming more cautious the closer it gets, enhancing the guidance of local search and improving convergence speed.

[0034] 3) Balance parameter R Population information exchange is introduced into the equilibrium parameter R, and the specific formula is as follows: (6); In the formula, This represents the current average position of the population. Weighting for information exchange; like If the sand cat enters the exploration phase, its next location can be a random combination of its prey and its current location, allowing for a global search; if The Sand Cat will enter the attack and exploitation phase, conducting localized searches.

[0035] In the traditional Sand Cat algorithm, the balance parameter during the exploration and exploitation phases... Its formula is: Compared to the traditional Dune Cat swarm algorithm, this invention utilizes overall population information to guide the search direction, enhances population cooperation, and avoids getting trapped in local optima.

[0036] 4) Elite guidance and disturbance strategy This strategy is a pair of complementary core mechanisms designed to accelerate algorithm convergence through "elite guidance" while maintaining population diversity through "adaptive perturbation," thereby effectively balancing the algorithm's development and exploration capabilities and avoiding premature convergence.

[0037] ① Elite guidance mechanism Elite guidance is achieved by guiding ordinary individuals in the population towards the current elite individuals (i.e., the current global optimal solution). By aligning the location information of the population, the population can be guided to quickly focus on potentially better regions, thereby improving the convergence speed. The specific method is as follows: The elite guidance item has been incorporated into the position update formula during the prey attack phase. The modified formula is as follows: ; In the formula: This represents the global historical best position (elite individual) in the t-th iteration. The elite guiding factor is a constant, usually set to 0.1~0.3.

[0038] New items This forms a directional vector pointing towards elite individuals, causing the current individual to be additionally pulled by elite individuals on top of the original update mechanism.

[0039] ② Adaptive perturbation mechanism The perturbation strategy injects diversity into the population by applying controlled random perturbations to elite individuals, preventing premature convergence caused by excessive clustering of all individuals in a local optimum. The specific method is as follows: When the algorithm detects the global optimal solution after K consecutive iterations (e.g., K=5) The improvement amount is less than the preset threshold At that time, it was determined that the situation had stalled, and then... Apply a Gaussian perturbation, and the new position after the perturbation. The calculation is as follows: ; In the formula: It is a random number that follows a standard normal distribution, providing the direction and magnitude of the random disturbance. It is the dynamically decaying disturbance intensity coefficient, and its calculation formula is: ,in This represents the initial strength. This design allows the perturbation to be stronger in the early stages of iteration to promote exploration, and weaker in the later stages of iteration to facilitate fine-tuning. It represents the range of the search space in each dimension, and is used to scale the perturbation amplitude to adapt it to the problem scale.

[0040] Effect: 1. Elite-guided learning utilizes historical best information, which significantly accelerates the algorithm's convergence process and enhances local exploitation capabilities.

[0041] 2. Adaptive perturbation introduces randomness in a targeted manner when the algorithm stagnates, effectively helping the population escape the trap of local optima and enhancing global exploration capabilities.

[0042] 3. The synergy between the two enables the improved SCSO algorithm to achieve a better dynamic balance between "utilizing known good regions" and "exploring unknown regions", thereby improving the overall optimization performance and robustness of the algorithm.

[0043] Step 3: Construct an XGBoost-LightGBM-LSTM fusion prediction model, which is trained using a runway deviation risk dataset for time-series data prediction. XGBoost Model: The prediction calculation formula of the XGBoost model is expressed as follows: (7); In the formula: This represents the prediction result for sample i after the t-th iteration; To sample The prediction result output by the k-th tree after inputting the data; This represents the prediction result for the (t-1)th tree; The model prediction result for the t-th tree; The objective function of the XGBoost model is expressed as follows: (8); In the formula: The objective function of the model; The corresponding loss function; This is a regularization term.

[0044] XGBoost Model: The core design idea of ​​the XGBoost model lies in constructing a series of learning models with relatively limited performance. These models are not trained independently, but rather exhibit a progressive learning relationship. Each subsequent model uses the biases generated by the preceding model during the prediction process (i.e., the difference between the actual result and the predicted result) as a new learning target for targeted training, continuously correcting the shortcomings of the earlier models to improve overall performance. Finally, the outputs of all these carefully designed base models are integrated and superimposed to achieve a prediction effect that is superior to that of a single model.

[0045] LightGBM Model The LightGBM model, proposed by Microsoft in 2015, is a new boosting framework that introduces two new technologies on top of the traditional GBDT: gradient-based one-side sampling (GOSS) and exclusive feature bundling (EFB). GOSS eliminates a large portion of data with very low gradients, using only the remaining data to estimate information gain, thus avoiding the influence of the long tail of low-gradient data. Since data with large gradients are more important for information gain, GOSS can still achieve fairly accurate estimates with significantly less data than traditional GBDT. EFB bundles mutually exclusive features to reduce the number of features. Furthermore, in the traditional GBDT algorithm, the most time-consuming step is to enumerate all possible feature points on the sorted feature values ​​using a pre-sorted approach and then find the optimal split point. LightGBM replaces the traditional pre-sorted approach with a histogram algorithm to reduce memory consumption.

[0046] LSTM Model: The LSTM (Long Short-Term Memory) model, as a key improvement on traditional recurrent neural networks, achieves its core breakthrough by designing a more sophisticated internal unit architecture, such as... Figure 2 As shown. This special structure enables it to effectively address the common problems of gradient vanishing and gradient amplification during training of conventional recurrent neural networks. The specific formula is as follows: (9); (10); (11); (12); (13); (14); In the formula, , They are time Corresponding input and hidden states; , and These represent the states of the forget gate, input gate, and output gate, respectively. , and These represent the states of the neuron to be updated and the unit state, respectively. , , , and , , , These represent the weight matrix and bias parameters of each gate, respectively; This represents the sigmoid activation function used in the hidden layer. Figure 2 middle, This indicates the hidden state at the current time and the previous time step. This indicates the generated prediction result.

[0047] The LSTM model consists of three parts: an input gate, a forget gate, and an output gate. The LSTM model is used to obtain long-term dependency information of runway risk time series data. The LightGBM model includes a gradient one-sided sampling module and an independent feature merging module. The gradient one-sided sampling module is used to remove data with very small gradients and perform information gain on the remaining data. The independent feature merging module is used to bundle mutually exclusive features into one feature and reduce the number of features. The XGBoost model constructs several weak learners. Each weak learner uses the fitting error of the previous weak learner as the learning target to continue fitting. Then, all weak learners are accumulated and the prediction result is output.

[0048] Step four: Optimize the fusion prediction model using the improved Sand Cat Optimization Algorithm (ISCSO); The specific optimization methods are as follows: Step A: The fusion prediction model obtains the time series dataset of risk data from the runway deviation risk quantification model. The specific method is as follows: 1) Retrieve runway overrun incident data from the Civil Aviation Safety Information Network from January 2017 to August 2025, and use the monthly risk value formula to calculate the monthly risk value for this aircraft type over 104 months. ; In the formula: i represents the "runaway" event in each month, from 1 to n; 2) To ensure that the diversity of different feature data value ranges and units does not adversely affect model training, the monthly average risk values ​​for 104 months were normalized using the following formula: (12); in , It is the number of eigenvalues; , These are the values ​​before and after normalization, respectively. , The first The minimum and maximum values ​​of the feature before normalization.

[0049] Step B: Initialize the sand cat population X using the improved ISCSO algorithm, set the maximum number of iterations T, nonlinear adjustment factor α, maximum perception distance D, information exchange weight γ, and related parameters of the elite guidance strategy, and optimize the XGBoost-LightGBM-LSTM fusion prediction model. (1) The parameters to be optimized include: XGBoost’s n_estimators, max_depth, learning_rate, subsample; LightGBM’s num_leaves, min_data_in_leaf, feature_fraction; LSTM units, dropout_rate, learning_rate; In step C, each sand cat performs exploration and development behaviors and updates its location based on the improved ISCSO algorithm, which combines four mechanisms: nonlinear adaptive parameters, dynamic perspective simulation, population information exchange, elite guidance and perturbation. Step D: After each iteration, calculate the RMSE value of the XGBoost-LightGBM-LSTM fusion model corresponding to the parameter combination represented by each sand cat, and update the global optimal solution xbest and the corresponding optimal parameter combination. Step E: Determine if the maximum number of iterations T has been reached: if so, output the optimal parameter combination; otherwise, return to continue iterating. Step F uses the fitness function RMSE as an evaluation metric to guide the ISCSO algorithm in finding the optimal parameter combination in the search space. The fitness function is used to evaluate the quality of each sand cat location in SCSO. This paper uses the RMSE (root mean square error) of the test set as the fitness value, and the calculation formula is as follows: (11); In the formula, The number of samples; The actual value; This is a predicted value; The fitness function evaluates the quality of each set of parameters, guiding the ISCSO algorithm to find the optimal parameter combination in the search space. Through continuous iteration, the SCSO algorithm gradually finds the parameter combination that minimizes the RMSE, thereby optimizing the model's predictive performance. Step G: After optimization, the optimal parameters obtained in step F are assigned to the XGBoost, LightGBM, and LSTM models respectively to complete the optimization of the runway deviation risk prediction model.

[0050] Step 5: Evaluate and test the runway deviation risk prediction model. Use test set data to test the optimized XGBoost-LightGBM-LSTM model, calculate the RMSE, MAE and other indicators of the prediction results, evaluate the model performance and draw a comparison chart between the prediction curve and the actual value.

[0051] The above provides a detailed description of the runway deviation risk prediction model based on the improved SCSO algorithm provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Modifications and improvements to this invention are possible without exceeding the concept and scope specified in the appended claims. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. An improved SCSO algorithm-based risk prediction model for drift-off-runway, characterized in that: The off-runway risk prediction model mainly consists of an off-runway risk quantification model and an XGBoost-LightGBM-LSTM fusion prediction model, and the specific construction method is as follows, Step one, construct the off-runway risk quantification model. The model identifies the core parameters directly related to the off-runway risk from the QAR data source through event analysis, industry reports and expert experience, and quantifies the parameters by scoring and assigning values. The total risk index TRI obtained by adding the scores of all parameters identified is used to collect the off-runway risk of each time period in chronological order to obtain the off-runway risk dataset; Step two, improve the sand cat optimization algorithm mechanism. In the search prey stage, a nonlinear adaptive sensitivity parameter is introduced to replace the original linear decreasing mechanism. In the attack prey stage, a dynamic perspective simulation mechanism is introduced to simulate the visual perception of the sand cat. A population information exchange mechanism is introduced in the balance parameter to enhance the local development capability. Elite guidance and disturbance strategies are introduced to avoid premature convergence; Step three, construct a fusion prediction model containing LSTM model, LightGBM model and XGBoost model. The fusion prediction model is trained for time series data prediction using the off-runway risk dataset; Step four, optimize the fusion prediction model by the improved sand cat optimization algorithm; Step five, evaluate and test the off-runway risk prediction model.

2. The off-runway risk prediction model based on the improved SCSO algorithm according to claim 1, wherein: In step one, the specific quantification method of the off-runway risk quantification model is as follows: First, collect and process QAR data: Second, assign different risk index scores to each parameter according to different states. The specific method is as follows: Classify the state types of key parameters. According to the rules of the risk quantification table, safety experts or trained analysts determine the state of each parameter and select the corresponding risk index score. In cases where it is difficult to determine, the conservative principle should be followed or submitted to the expert group for review; Then, add the scores corresponding to the specific states of each parameter involved in the event to obtain the total risk index TRI of the event; Finally, to better distinguish and manage off-runway, set the off-runway risk level threshold, which is divided into four levels: red, orange, yellow and blue.

3. The off- center runway excursion risk prediction model based on the improved SCSO algorithm according to claim 2, characterized in that: The specific method for collecting and processing QAR data is as follows, Step a. Event identification: identify possible off-runway events through QAR over-limit event reports, crew reports and ATC reports; Step b. Data extraction: extract radio altitude, vertical speed, air-ground signal, acceleration, throttle position, brake pressure, reverse thrust state and related parameters from QAR data; Step c. Multi-source information fusion: integrate weather data, NOTAM, runway condition reports, ATC call recordings and cockpit audio recordings to accurately determine runway conditions, weather conditions, crew decisions and control interactions.

4. The off- center runway excursion risk prediction model based on the improved SCSO algorithm according to claim 3, characterized in that: The off-runway risk level is divided as follows, TRI >= 90, extremely high risk, indicating a high probability of accident; Orange level: 75 <= TRI < 90, high risk, which needs to be taken seriously; Yellow level: 60≤TRI<75, medium risk, as a safety warning information, for daily risk monitoring and trend analysis; Blue level: TRI<60, low risk, as a safety warning information, for daily risk monitoring and trend analysis.

5. The off- center runway excursion risk prediction model based on the improved SCSO algorithm according to claim 1, wherein: In step two, the improved sand cat optimization algorithm is as follows: 1) Search for prey In the search for prey stage, the sand cat will find the best prey in its search range and move to the prey position. The position update is shown in the following formula: （1）； wherein is the initial position of the cth dune cat at the tth iteration; is the sensitivity of the dune cat, is the optimal individual of the tth generation; is a random number between 0 and 1. wherein the sensitivity of the dune cat is given by the formula ， （2）； wherein is a non-linear adaptive sensitivity parameter, is a random number between 0 and 1; Nonlinear adaptive sensitivity parameter The formula is as follows （3）； In the formula, is the sand dune cat hearing characteristics, the value is a constant, usually set to 2; T is the maximum number of iterations; is a nonlinear adjustment factor, usually 1.5-2.5; 2) Attack prey In this stage, each sand cat uses a dynamic perspective to approach the hunting position along different circular directions in space. The position update formula is shown in the following formula: （4）； In the formula: is a random distance; is a dynamic view angle; is a random number, which is randomly disturbed to ensure that the algorithm has a certain randomness and helps to avoid falling into a local optimal solution; is the current position of the cth dune cat at the tth iteration; wherein the dynamic viewing angle is given by the formula: d （5）； wherein is the maximum perceived distance, is the current distance optimal solution, d represents a random number uniformly distributed in the range [0, 1). 3) Balance parameter R Introduce population information exchange in the balance parameter R. The specific formula is as follows （6）； wherein is the current population average position, is the information exchange weight; If , the dune cat enters the exploration stage, and the next position can be at a random position of the prey and the current position, for global search; if , the dune cat will enter the attack and utilization stage, for local search; 4) Elite guidance and disturbance strategy Through "elite guidance", the algorithm convergence is accelerated, and through "adaptive disturbance", the population diversity is maintained, so as to effectively balance the development and exploration ability of the algorithm and avoid premature convergence.

6. The off-runway risk prediction model based on the improved SCSO algorithm according to claim 5, characterized in that: The elite guiding mechanism is to guide the population to focus on the potential better area quickly by letting the ordinary individuals in the population move to the position of the contemporary elite individual, i.e. the current global optimal solution The position information of the contemporary elite individual, thereby improving the convergence speed, and the specific method is as follows: In the position update formula of the attack prey stage, the elite guidance term is integrated, and the modified formula is: ； In the formula, is the global historical optimal position in the tth iteration, i.e., the elite individual; is an elite guide factor, which is a constant, usually set to 0.1-0.3; The disturbance strategy is to apply controllable random disturbance to the elite individuals to inject diversity into the population and prevent premature convergence caused by excessive aggregation of all individuals in a local optimal region. The specific method is as follows: When the algorithm detects that the improvement of the global optimal solution is less than a preset threshold value for K consecutive iterations , it is determined that it is stuck, and a Gaussian disturbance is immediately applied to the new position after the disturbance is calculated as follows:​ ； wherein: is a random number subject to standard normal distribution, providing the random perturbation direction and amplitude; is a dynamic decayed perturbation intensity coefficient, whose formula is wherein is the initial intensity, characterizes the range of the search space in each dimension, used to scale the perturbation amplitude to make it compatible with the problem scale.

7. The off- center runway excursion risk prediction model based on the improved SCSO algorithm according to claim 1, wherein: In step four, the improved sand cat optimization algorithm optimizes the fusion prediction model as follows, Step A, based on the off-runway risk quantification model, obtain the risk data set of a certain type of aircraft in a corresponding time period, and normalize the risk data; Step B, initialize the sand cat population X of the improved ISCSO algorithm, set the maximum iteration number T, the nonlinear adjustment factor α, the maximum perception distance D, the information exchange weight γ, and the related parameters of the elite guidance strategy, and optimize the XGBoost-LightGBM-LSTM fusion prediction model; Step B, each sand cat performs exploration and development behavior according to the improved ISCSO algorithm, combined with the introduction of nonlinear adaptive parameters, dynamic perspective simulation, population information exchange, and elite guidance and disturbance of four mechanisms, to realize position update; Step D, after each iteration, calculate the RMSE value of the XGBoost-LightGBM-LSTM fusion model corresponding to the parameter combination represented by each sand cat, and update the global optimal solution xbest and the corresponding optimal parameter combination; Step E, determine whether the maximum iteration number T is reached: if yes, output the optimal parameter combination; otherwise, return to continue iteration; Step F, use the fitness function RMSE as the evaluation index to guide the ISCSO algorithm to find the optimal parameter combination in the search space; Step G, after optimization, assign the optimal parameters obtained in step F to XGBoost, LightGBM and LSTM models respectively to complete the optimization of the off-runway risk prediction model.

8. The improved SCSO algorithm-based risk prediction model for drift-off- runway according to claim 7, characterized in that: In step A, the specific method of obtaining risk data is as follows: 1) Call the off-runway unsafe event data of the corresponding month of a certain model, get the risk data through the off-runway risk quantification model, and use the monthly risk value formula to obtain the monthly risk value of the model, the monthly risk value formula is as follows, ； In the formula: i represents the "off-runway" event in each month, from 1 to n; 2) In order to ensure that the diversity of the value range and unit of different characteristic data will not have adverse effects on model training, the corresponding monthly average risk value is normalized, and the normalization uses the following formula: （12）； wherein , is the number of eigenvalues; , are the values before and after normalization, respectively; , are the minimum and maximum values of the i-th feature before normalization, respectively. are the minimum and maximum values of the i-th feature before normalization, respectively.

9. The improved SCSO algorithm-based off-pavement and runway excursion risk prediction model according to claim 7, characterized in that: In step E, the evaluation index uses the RMSE root mean square error of the test set as the fitness value, and the calculation formula is as follows: （11）； In the formula, is the number of samples; is the true value; is the predicted value.

10. The off- center runway excursion risk prediction model based on the improved SCSO algorithm according to claim 1, wherein: In step five, the test set data is used to test the optimized XGBoost-LightGBM-LSTM model, the RMSE and MAE indexes of the prediction result are calculated, the model performance is evaluated, and the prediction curve and the real value comparison graph are drawn.

Images