Multi-model fusion-based operation management risk prediction method for converting small divergence into underrun
By building a hierarchical safety risk indicator system and improving the Black Kite optimization algorithm, combined with the temporal convolutional network and bidirectional LSTM model, the problems of data dispersion and information fragmentation among small and scattered airlines have been solved, high-precision operational risk prediction has been achieved, and a dynamic risk management tool has been provided.
Patent Information
- Application Number
- CN202510789843.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-13
- Publication Date
- 2025-09-26
AI Technical Summary
Existing technologies are difficult to effectively adapt to the actual needs of small and scattered airlines with scattered data and fragmented information, resulting in redundant risk prediction model indicators, high computational complexity, and limited model generalization capabilities, making it difficult to accurately predict operational risks.
A hierarchical security risk indicator system is constructed, combining the temporal convolutional network and the bidirectional LSTM model. The improved Black Kite optimization algorithm is used to optimize the model parameters to capture the trend, seasonality and nonlinear relationship characteristics of the data. The model parameters are optimized through the improved Black Kite optimization algorithm model to predict the operational risk values of several consecutive time span groups.
It improves the accuracy of risk prediction and the generalization ability of the model, and can accurately predict the operational risk values of several consecutive time span groups, provide dynamic risk management technical support, and provide important reference for the safety decision-making and risk prevention and control strategy formulation of small and scattered airlines.
Smart Images
Figure CN120706628A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of risk management of small and scattered airlines that have become delinquent airlines, and in particular to a method for predicting operational management risks of small and scattered airlines that have become delinquent airlines based on a fusion of multiple models. Background Art
[0002] The concept of "small, scattered, transfer-debt carriers" is widely discussed in civil aviation safety regulation, particularly in the oversight of main-base transport airlines. Due to these characteristics, these airlines often face higher safety risks and regulatory challenges. For example, to address these characteristics, innovative regulatory mechanisms have been developed to enhance regulatory effectiveness and ensure the safe operation of airlines. In recent years, the civil aviation industry has continued to develop rapidly, with the transport aviation market expanding and the number of airlines increasing. Against the backdrop of intensified market competition, some airlines have been categorized as "small, scattered, transfer-debt carriers" (i.e., "small scale, scattered operations, changes in ownership, transfers to other bases, and insufficient investment") due to their small size, decentralized operations, frequent equity changes, frequent base relocations, and insufficient resource investment. These airlines face unique safety management risks and challenges, such as unstable organizational structures, uneven allocation of safety resources, and decentralized operational processes. Their safety management capabilities directly impact the safety of passengers and property and the overall safety level of the industry.
[0003] The core of civil aviation safety management lies in risk prediction and prevention. Existing research shows that risk modeling based on historical unsafe incident data is an important means of identifying potential hazards and predicting risk trends. Domestic and international scholars have explored risk prediction models using various methods. Meng Qingsen et al. employed spatiotemporal threshold analysis to construct a Bayesian optimized random forest model to predict secondary highway accidents. Yu Sixuan et al. used a sparse denoising autoencoder algorithm to accurately predict the flight rate per 10,000 passengers. Sun Ruishan et al. used improved correlation analysis to identify key factors influencing civil aviation accident symptoms and then used triangular fuzzy numbers to predict them from time series. Yuan Liangyun et al. constructed a civil aviation safety risk indicator system based on text mining and the Analytic Hierarchy Process (AHP) and proposed a hierarchical risk response strategy. Wang Jian et al. proposed a data-driven general aviation risk assessment system, promoting a shift in regulatory models from post-event reactive management to data-driven approaches. Wu Liangchuan et al. designed and implemented an airline risk assessment and early warning system, integrating multiple functions to improve risk management efficiency. Xiong Haitao et al. used a gray prediction method to analyze the risk coefficient of unsafe incidents for airlines and, combined with a risk matrix, effectively identified potential triggers for extreme accidents. Xu Yi et al. constructed a principal component analysis (PCA)-genetic algorithm (GA)-back propagation (BP) combined model to improve the accuracy of civil aviation accident incident rate prediction. Wang Yantao et al. reconstructed the phase space based on multivariate chaotic time series and used the radial basis function (RBF) model to predict flight operation risks, verifying the effectiveness of chaos theory in short-term risk prediction. Zhao Shan et al. designed a bidirectional long short-term memory network (BiLSTM)-back propagation (BP) safety risk prediction model, optimized the model parameters, and significantly improved the accuracy, providing a new method for dynamic risk warning in civil aviation. Although these studies provide theoretical and technical support for risk prediction, they generally rely on large-scale multidimensional data and have problems such as redundant indicators, high computational complexity, and limited model generalization ability. They are difficult to adapt to the actual needs of "small, scattered, and poorly-funded" airlines with scattered data and fragmented information. Summary of the Invention
[0004] The purpose of the present invention is to provide a method for predicting the operational management risks of small, scattered, and under-performing airlines based on the fusion of multiple models. The airline operational risk prediction combined model can capture the trend, seasonality, time series characteristics and nonlinear relationship characteristics of the data, and use the improved black kite optimization algorithm model to optimize the model parameters, which can accurately predict the operational risk values of several consecutive time span groups.
[0005] The purpose of the present invention is achieved through the following technical solutions:
[0006] A method for predicting operational management risks of small, scattered, and under-performing airlines based on a fusion of multiple models, comprising: S1, constructing a safety risk indicator system according to a hierarchical architecture, obtaining sample data of historical events before a research prediction date, dividing the data into several time span groups according to a natural year, wherein the sample data of historical events includes data corresponding to the safety risk indicator system and operational risk value labels, calculating the operational risk value of each time span group, and sequentially aggregating the operational risk values of each time span group according to a time series to form an operational risk data set;
[0007] S2. Construct an airline operation risk prediction combination model that includes a temporal convolutional network and a BiLSTM model, input the operation risk dataset into the airline operation risk prediction combination model for model training; construct an improved Black Kite optimization algorithm model to optimize the model parameters of the airline operation risk prediction combination model;
[0008] S3. Collect data related to the safety risk indicator system on the forecast date and input it into the airline operation risk forecast combination model to output the operation risk values for several consecutive time span groups after the forecast date.
[0009] In order to better implement the present invention, the temporal convolutional network of the airline operation risk prediction combined model is used to perform dependency capture convolution coding processing on the historical event sample data associated with the safety risk indicator system in the operation risk data set according to the time series; the BiLSTM model is composed of a forward LSTM unit and a backward LSTM unit and performs bidirectional processing to obtain feature information. Both the forward LSTM unit and the backward LSTM unit adopt a gating mechanism in which a forget gate, an input gate and an output gate cooperate with each other.
[0010] Preferably, the improved black kite optimization algorithm model randomly sets the black kite position as a potential solution in the search space to perform global search and local search, iteratively updates the black kite position to explore the optimal solution of the model parameters, and the improved black kite optimization algorithm model adopts a time-series adaptive attack strategy to dynamically adjust the attack step coefficient w. The expression for updating the individual position of a black kite is: in Indicates that the i-th black-winged kite updates the position of the j-th dimension in the t+1-th iteration, represents the position of the i-th black kite in the j-th dimension at the t-th iteration, t is the current iteration number, maxt is the maximum iteration number, p is a preset constant, r is a random number in [0, 1], and n represents the Cauchy mutation result.
[0011] Preferably, the dimension of the improved black kite optimization algorithm model adopts a dimension equal division linear locking strategy, divides each dimension d evenly into n1 equal segments, calculates the fitness value of each equal division point, explores the potential optimal solution on the dimension by traversing the equal division points on each dimension, and selects the optimal equal division point to construct a new individual:
[0012] Among them, ub(d) and lb(d) are the upper and lower bounds of the dimension, respectively, k∈[0,n1]; by changing the value of k, different equal division points can be obtained, and the individual fitness values corresponding to these equal division points can be calculated. The point with the best fitness value is selected as the representative of the dimension.
[0013] Preferably, the improved black kite optimization algorithm model introduces an improved differential evolution strategy, and the differential evolution strategy designs a dynamic scaling factor F, which is expressed as follows: The improved black kite optimization algorithm model maintains a large mutation amplitude in the early stage of the algorithm to enhance the global search ability, and gradually reduces the mutation intensity as the iteration deepens to achieve a fine search of the solution space; in the mutation operation, the elite individual is combined to generate the mutation vector v i，g , the expression is: i，g =x r1，g +F·(x r2，g -x r3，g ), where x r1，g 、x r2，g 、x r3，g To randomly select three different individuals from the population; retain the dominant gene u through binomial crossover i，j，g , the expression is as follows:
[0014] Where CR is the crossover probability, j rand is a randomly generated integer. This operation ensures that the mutation vector v i，j，g At least one dimension of information is retained, while retaining the original individual x i，j，g The new individual not only contains the information of the mutation, but also retains the dominant genes of the original individual.
[0015] Preferably, when the improved black kite optimization algorithm model detects that the fitness fluctuation is lower than the threshold δ, part of the population is reinitialized, and the expression is as follows:
[0016] x drs =x+r·(r·(ub+lb-x)-x), where x is the current individual, ub and lb are the upper and lower bounds of the individual value, and r is a random number in [0, 1]. By perturbing the current individual x, a new individual x is generated. drs ; Retain high-quality individuals through competitive selection, the expression is as follows: where f(x drs ) is used to generate a new individual x drs The fitness of f(x r ) is a randomly initialized individual x r The fitness of x r To randomly initialize individuals, by comparing x drs and x r fitness and select individuals with better fitness to ensure that high-quality individuals are always retained in the population.
[0017] Preferably, the safety risk index system includes four dimensional levels, namely, event level, event type, incident stage and event cause. The event level includes several element levels including general events and general event accident signs. The event type includes several element levels including loss of control / stall, wind shear, bird strike, system failure / malfunction / jamming. The incident stage includes several element levels including taxiing, takeoff, climb, cruise and landing. The event cause includes several element levels including air traffic control, maintenance, weather accidents and ground support.
[0018] Preferably, the safety risk index system includes four dimensional levels, namely, event level, event type, incident stage and event cause, and weights are determined by the Delphi method respectively. The element levels under the four dimensional levels are also respectively set with weights.
[0019] Preferably, the improved black kite optimization algorithm model selects a population leader to migrate the population. When the population fitness value under the population leader is lower than the threshold, the population leader will give up leadership and join the migrating population, and use Cauchy mutation to obtain the mutated population leader; otherwise, the population leader will lead the population to migrate, and the position expression of the mutated leader is as follows: where X leader For the leader before mutation, is the leader after mutation, γ is the set Cauchy distribution parameter, and Cauchy(0, 1) is a random number that obeys the standard Cauchy distribution.
[0020] Preferably, the time span groups are divided by day, and the time span of the time span groups is N1 days; or the time span groups are divided by month, and the time span of the time span groups is one natural month.
[0021] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0022] (1) The present invention constructs a safety risk indicator system with a hierarchical architecture, obtains sample data of historical events before the research prediction date, divides them into several time span groups according to a natural year, counts the operational risk values of each time span group, and aggregates the operational risk values of each time span group in sequence according to the time series to form an operational risk data set; the airline operational risk prediction combination model can capture the trend, seasonality, time series characteristics and nonlinear relationship characteristics of the data, and uses the improved black kite optimization algorithm model to optimize the model parameters, which can accurately predict the operational risk values of several consecutive time span groups.
[0023] (2) The present invention optimizes the weight of indicators by combining the frequency of event occurrence, stage sensitivity and potential harmfulness, and shows significant advantages in processing nonlinear and dynamic risk data; by optimizing time series feature extraction and dependency modeling through lightweight intelligent algorithms, the model improves prediction accuracy, provides dynamic risk management technical support for resource-constrained "small, scattered, and shifting to under-resourced" airlines, and provides important technical reference for the formulation of safety decisions and risk prevention and control strategies.
[0024] (3) The improved black kite optimization algorithm model of the present invention shows significant advantages in hyperparameter optimization by introducing an adaptive attack strategy, a linear locking strategy and a dynamic response mechanism; the adaptive attack strategy balances global search and local mining capabilities to accelerate model convergence; the linear locking strategy optimizes the search efficiency of discrete parameters through a dimension equal division and selection mechanism; the dynamic response mechanism quickly adapts to changes in data distribution to prevent the algorithm from falling into a local optimal solution. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 This is a flow chart of the method for predicting the operational management risks of small and scattered airlines that are in arrears according to the present invention;
[0026] Figure 2 Schematic diagram of the method principle of the MBKA-TCN-BiLSTM hybrid model in the embodiment;
[0027] Figure 3 Schematic diagram of the method principle for improving the black kite optimization algorithm model in the embodiment;
[0028] Figure 4 This is a time series statistical chart of operational risk values obtained by extracting part of the data in the embodiment;
[0029] Figure 5 This is a graph showing the loss function calculation results of the MBKA-TCN-BiLSTM hybrid model in the embodiment. DETAILED DESCRIPTION
[0030] Below in conjunction with embodiment, the present invention is described in further detail:
[0031] Example
[0032] like Figure 1 As shown in the figure, a method for predicting the operational management risk of small and scattered airlines based on the fusion of multiple models is provided, which includes:
[0033] S1. Construct a security risk indicator system based on a hierarchical structure and obtain sample data of historical events before the research prediction date (this embodiment uses four years of historical data, see Figure 4 (The data from the four-year historical data is intercepted to conduct time-series statistical mapping of operational risk values.) A number of time span groups are divided according to a natural year. The historical event sample data includes data corresponding to the safety risk indicator system and operational risk value labels. The operational risk values of each time span group are statistically analyzed, and the operational risk values of each time span group are sequentially aggregated according to the time series to form an operational risk data set. In some embodiments, the safety risk indicator system includes four dimensional levels: event level, event type, incident stage, and event cause. The event level includes several element levels, including general events and general event accident signs. The event type includes several element levels, including loss of control / stall, wind shear, bird strike, system failure / malfunction / jam. The incident stage includes several element levels, including taxiing, takeoff, climb, cruise, and landing. The event cause includes several element levels, including air traffic control, maintenance, weather accidents, and ground support. The safety risk index system includes four dimensions: event level, event type, incident stage and event cause. The weights of these four dimensions are determined by the Delphi method. The four dimensions are W g 、W t 、W p 、W c (i.e., event level weight W g , event type weight W t , weight W at the incident stage p , the cause of the incident W c), the operational risk value of the time span group is obtained by weighted evaluation according to the data of the above four dimensional levels. Preferably, the element levels under the four dimensional levels are also set with weights. For example, the weight of the cause of the event is assigned according to its frequency of occurrence and potential danger; for example, a mechanical failure may be given a higher weight due to its high risk of directly causing an accident. The weight of the incident stage, the take-off and landing stages are generally considered to be the most dangerous stages in the flight process, so the weight of the relevant risk events in these stages should be appropriately increased; in contrast, although the cruising stage is usually relatively safe, risks such as flying too low and yaw may still occur; therefore, the risk weights of different flight stages need to be dynamically adjusted to ensure that the prediction model can reflect the global changes in risks. The historical event sample data is analyzed according to the time series, which can reveal the potential high-risk periods and fluctuation patterns; based on these time series data, a risk fluctuation airline operation risk prediction combination model can be constructed to predict future risk changes and help airlines take prevention and control measures in advance. For example, in the event type, the weight of controlled flight into terrain / obstacle is 0.1, the weight of missing / worn / damaged parts is 0.04, the weight of wind shear is 0.08, the weight of bird strike is 0.05, the weight of system failure / malfunction / jam is 0.09, the weight of engine shutdown is 0.05, etc.; in the event cause, the weight of mechanical is 0.15, the weight of weather accident is 0.15, etc.; in the incident phase, the weight of climb is 0.15, the weight of takeoff is 0.2, the weight of cruise is 0.1, etc. In this embodiment, the sum of all element levels can be normalized to 1 and the normalization of each weight can be performed.
[0034] In some embodiments, the time span group is divided by day, and the time span of the time span group is N1 days; or the time span group is divided by month, and the time span of the time span group is one natural month. Example: A natural year contains 12 natural months, and one natural month can be divided into 6 time span groups (i.e., 5 days of time span are divided into one time span group, such as Figure 4 As shown, a time span group is divided according to a 5-day time span for aggregation of operational risk values); of course, the time span is divided according to natural months. A natural year contains 12 natural months, and a natural year is divided into 12 time span groups.
[0035] S2. Construct an airline operation risk prediction combination model that includes a temporal convolutional network (TCN) and a BiLSTM model, and input the operation risk dataset into the airline operation risk prediction combination model for model training. Figure 2As shown, an improved black kite optimization algorithm model is constructed to optimize the model parameters of the airline operation risk prediction combination model. The fusion multi-model of the present invention includes an improved black kite optimization algorithm model (abbreviated as MBKA model, MBKA is the English abbreviation of the improved black kite optimization algorithm) and an airline operation risk prediction combination model (also known as TCN-BiLSTM hybrid model). Therefore, the fusion multi-model of the present invention is also called MBKA-TCN-BiLSTM hybrid model (such as Figure 2 As shown). The temporal convolutional network (also known as the TCN model, which uses a dilation factor of d=2 in the dilated convolution layer to expand the receptive field layer by layer, thereby capturing global features across time steps. The activation function is ReLU.) of the airline operation risk prediction combined model is used to perform dependency capture convolution coding processing on the historical event sample data associated with the safety risk indicator system in the operation risk dataset according to the time series. The BiLSTM model consists of a forward LSTM unit and a backward LSTM unit and processes bidirectionally to obtain feature information. Both the forward LSTM unit and the backward LSTM unit use a gating mechanism in which a forget gate, an input gate, and an output gate cooperate with each other. The BiLSTM model in this embodiment contains 256 hidden units, an input dimension of 128, and an output dimension of 512 (forward and backward are 256-dimensional splicing). An example is as follows: If the input time series data is x=[x1, x2,…, x T ], the hidden state of the forward LSTM at time t and memory units The calculation is as follows:
[0036] Forget Gate:
[0037] Input Gate:
[0038] Candidate memory cells:
[0039] Memory unit update:
[0040] Output Gate:
[0041] Hide status update:
[0042] Among them, W f , W i , W c , W o is the weight matrix, b f , b i , b c , b o is the bias vector, σ is the sigmoid activation function, ⊙ represents element-wise multiplication, and tanh is the hyperbolic tangent activation function.
[0043] Hidden state of the backward LSTM unit at time t and memory units The calculation is similar, but the processing order is reversed. The gating mechanism of the LSTM unit is the core of its ability to overcome the problem of long-term dependencies. The forget gate, input gate, and output gate work together to cleverly store important information in the memory unit, preventing information loss during long-term sequence processing. The BiLSTM model further expands the dimension of information acquisition through bidirectional processing.
[0044] In some embodiments, the improved black kite optimization algorithm model randomly sets the black kite position as a potential solution in the search space to perform global search and local search, iteratively updates the black kite position to explore the optimal solution of the model parameters, and the improved black kite optimization algorithm model adopts a time-series adaptive attack strategy to dynamically adjust the attack step coefficient w. In the early stage of the search, due to limited understanding of the location of the optimal solution, it is necessary to expand the population search range to quickly find possible high-quality areas and accelerate convergence. By increasing the attack step coefficient w, individuals can explore in a wider parameter space. As the search progresses, when approaching the optimal solution, it is necessary to enhance the local refined search capability to improve the accuracy of the solution. Therefore, in the later stage of the search, w is gradually reduced so that individuals can perform more detailed searches in the neighborhood of the optimal solution. The mathematical description expression of the attack step coefficient w is as follows: The expression for updating the individual position of a black kite is: in Indicates that the i-th black-winged kite updates the position of the j-th dimension in the t+1-th iteration, represents the position of the i-th black kite in the j-th dimension at the t-th iteration, where t is the current iteration number, maxt is the maximum iteration number, p is a preset constant, r is a random number in the range [0, 1], and n represents the Cauchy mutation result. The temporal adaptive attack strategy of this embodiment dynamically adjusts the search step size to quickly cover the parameter space in the early stages. By utilizing larger w and n values, individuals can explore more potential areas, accelerating convergence. Later, they focus on the neighborhood of the optimal solution. By reducing w and n values, individuals perform more precise searches near the optimal solution. This effectively coordinates the search intensity at different stages of the algorithm, improving search efficiency and convergence accuracy.
[0045] In some embodiments, the dimension of the improved Black Kite optimization algorithm model adopts a dimension equal division linear locking strategy, which evenly divides each dimension d into n1 equal segments, calculates the fitness value of each equal division point, and explores the potential optimal solution on the dimension by traversing the equal division points on each dimension, and selects the optimal equal division point to construct a new individual:
[0046] Where ub(d) and lb(d) are the upper and lower bounds of the dimension, respectively, and k∈[0,n1]. By varying the value of k, different equal-division points can be obtained. The individual fitness values corresponding to these equal-division points are calculated, and the point with the best fitness value is selected as the representative of that dimension. The linear optimization strategy for equal-division in this embodiment optimizes the local dimensions, fully utilizing the information of each dimension in the high-dimensional space. This prevents the algorithm from overlooking certain important potential solutions during the search process, helping the algorithm to find the global optimal solution more quickly.
[0047] In some embodiments, the improved Black Kite optimization algorithm model introduces an improved differential evolution strategy, which designs a dynamic scaling factor F, which is expressed as follows: The improved black kite optimization algorithm model maintains a large mutation amplitude in the early stage of the algorithm to enhance the global search ability, and gradually reduces the mutation intensity as the iteration deepens to achieve a fine search of the solution space. In the mutation operation, the elite individual is combined to generate the mutation vector v i，g , the expression is: i，g =x r1，g +F·(x r2，g -x r3，g ), where x r1，g 、x r2，g 、x r3，g To randomly select three different individuals from the population, retain the dominant gene u through binomial crossover i，j，g , the expression is as follows:
[0048] Where CR is the crossover probability, j rand is a randomly generated integer. This operation ensures that the mutation vector v i，j，g At least one dimension of information is retained, while retaining the original individual x i，j，g The differential evolution strategy of this embodiment not only maintains population diversity, but also controls the intensity of the mutation operation by dynamically adjusting the scaling factor F, thereby significantly improving the efficiency of global optimization.
[0049] In some embodiments, the improved black kite optimization algorithm model also adopts a dynamic response mechanism. When the improved black kite optimization algorithm model detects that the fitness fluctuation is lower than a threshold δ (in this embodiment, δ=0.01), part of the population is reinitialized. The expression is as follows:
[0050] x drs =x+r·(r·(ub+lb-x)-x), where x is the current individual, ub and lb are the upper and lower bounds of the individual value, and r is a random number in [0, 1]. By perturbing the current individual x, a new individual x is generated.drs . The high-quality individuals are retained through competitive selection, and the expression is as follows: where f(x drs ) is used to generate a new individual x drs The fitness of , where f() is the fitness function used in this embodiment, f(x r ) is a randomly initialized individual x r The fitness of x r To randomly initialize individuals, by comparing x drs and x r The fitness of the population is determined and individuals with better fitness are selected to ensure that high-quality individuals are always retained in the population. The dynamic response mechanism of this embodiment introduces reverse learning perturbations to quickly adapt to changes in data distribution while ensuring population diversity, thus preventing the algorithm from falling into a local optimal solution.
[0051] In some embodiments, the improved Black Kite optimization algorithm model selects a population leader to migrate the population. When the population fitness value under the population leader falls below a threshold, the population leader will give up leadership and join the migrating population. Cauchy mutation is used to obtain a mutated population leader. Otherwise, the population leader will lead the population to migrate. The position of the mutated leader is expressed as follows: where X leader For the leader before mutation, is the leader after mutation, γ is the set Cauchy distribution parameter, and Cauchy(0,1) is a random number following a standard Cauchy distribution. This helps the algorithm escape local optimal solutions and increases the likelihood of finding a better solution in the global search space. By continuously iteratively updating the position of the black kite and judging the quality of each solution based on the fitness function, the optimal solution to the problem is gradually explored.
[0052] The working principle of the improved black kite optimization algorithm model of the present invention is as follows Figure 3 As shown in the figure, the objective function and parameter configuration are first input, the initial population is generated through Cubic chaotic mapping and the individual fitness is calculated. Then, the adaptive attack strategy is executed to update the individual position. If the parameter is out of bounds, the search range is expanded in combination with the global worst solution. Otherwise, the dimension equal division strategy is adopted to screen high-quality new individuals. After the quality of the new individuals is judged, the differential evolution operation is driven by a dynamic scaling factor, and the fitness fluctuation is monitored in real time. When an anomaly is triggered, a reverse perturbation population is generated to compete again. Finally, the optimal parameters are output when the maximum number of iterations is reached. If the conditions are not met, the optimization cycle is continuously cyclically optimized, forming a complete "evaluation-attack-evolution-response" closed-loop logic chain to ensure efficient convergence to the global optimal solution.
[0053] S3. Collect data related to the safety risk indicator system on the forecast date and input it into the airline operation risk forecast combination model to output the operation risk values for several consecutive time span groups after the forecast date.
[0054] In order to verify the technical effect of the method of the present invention by integrating multiple models (i.e., MBKA-TCN-BiLSTM hybrid model), the data of aircraft unsafe events of a certain airline from a certain month in 2020 to 2024 from the Civil Aviation Safety Information Network were retrieved as data samples. The evaluation indicators of the MBKA-TCN-BiLSTM hybrid model of the present invention are Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (R-Square, R 2 ), the MBKA-TCN-BiLSTM hybrid model of the present invention has a mean absolute error MAE of 0.0146, a root mean square error RMSE of 0.0178, and a determination coefficient R 2 is 0.9153. If only the TCN-BiLSTM hybrid model is used to implement the method of the present invention, the mean absolute error MAE is 0.0179, the root mean square error RMSE is 0.0215, and the determination coefficient R 2 is 0.8751. After comparison, the MBKA-TCN-BiLSTM hybrid model used in this invention reduces the mean absolute error MAE by 0.33% and the root mean square error RMSE by 0.37% compared with the TCN-BiLSTM model. The determination coefficient R 2 The improvement was 4.02%, indicating that the improved black kite optimization algorithm of the present invention played a key role in optimizing the model, making the model prediction results more accurate and reducing the parameter adjustment process.
[0055] The loss function curve of the MBKA-TCN-BiLSTM hybrid model of the present invention is as follows Figure 5 As shown in the figure, it can be seen that the MBKA-TCN-BiLSTM hybrid model of the present invention has obvious advantages in all key indicators, which intuitively reflects that its prediction results are more accurate. It can effectively mine the complex features in the data and simplify the process of hyperparameter optimization, further verifying the prediction and evaluation advantages of the present invention. Figure 2As shown, the MBKA-TCN-BiLSTM hybrid model of the present invention combines the local feature extraction capabilities of the temporal convolutional network (TCN), the bidirectional time series modeling capabilities of BiLSTM, and the adaptive hyperparameter optimization strategy of MBKA. The temporal convolutional network (TCN) captures global risk patterns across time steps through dilated convolutions, and the BiLSTM model effectively constructs long-term dependencies in risk data. Combined with the dynamic optimization of key hyperparameters by the improved Black Kite optimization algorithm (MBKA), the model significantly improves the prediction accuracy and model generalization capabilities of complex time series risk data. By introducing an adaptive attack strategy, a linear lock optimization strategy, and a dynamic response mechanism, the improved Black Kite optimization algorithm (MBKA) exhibits significant advantages in hyperparameter optimization. The adaptive attack strategy balances global search and local mining capabilities to accelerate model convergence. The linear lock optimization strategy optimizes the search efficiency of discrete parameters through a dimension-equalizing optimization mechanism. The dynamic response mechanism quickly adapts to changes in data distribution, preventing the algorithm from falling into a local optimal solution. The MBKA-TCN-BiLSTM model of the present invention shows significant advantages in processing nonlinear and dynamic risk data; by optimizing time series feature extraction and dependency modeling through lightweight intelligent algorithms, the model improves prediction accuracy, provides a dynamic risk management tool for resource-constrained "small, scattered, and shifting to under-resourced" airlines, and provides an important reference for safety decision-making and the formulation of risk prevention and control strategies.
[0056] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for predicting operational management risks of small and scattered airlines based on the fusion of multiple models, characterized by: The methods include: S1. Construct a hierarchical safety risk indicator system. Obtain sample data of historical events before the research forecast date and divide them into several time span groups according to a natural year. The sample data of historical events includes data corresponding to the safety risk indicator system and operation risk value labels. Calculate the operation risk value of each time span group and aggregate the operation risk values of each time span group in sequence according to the time series to form an operation risk data set. S2. Construct an airline operation risk prediction combination model that includes a temporal convolutional network and a BiLSTM model, input the operation risk dataset into the airline operation risk prediction combination model for model training; construct an improved Black Kite optimization algorithm model to optimize the model parameters of the airline operation risk prediction combination model; S3. Collect data related to the safety risk indicator system on the forecast date and input it into the airline operation risk forecast combination model to output the operation risk values for several consecutive time span groups after the forecast date.
2. The method for predicting operational management risks of small and scattered airlines based on fusion of multiple models according to claim 1 is characterized by: The temporal convolutional network of the airline operation risk prediction combined model is used to perform dependency capture convolution coding processing on the historical event sample data associated with the safety risk indicator system in the operation risk dataset according to the time series; the BiLSTM model is composed of a forward LSTM unit and a backward LSTM unit and performs bidirectional processing to obtain feature information. Both the forward LSTM unit and the backward LSTM unit adopt a gating mechanism in which a forget gate, an input gate and an output gate cooperate with each other.
3. The method for predicting operational management risks of small and scattered airlines based on fusion of multiple models according to claim 1 is characterized by: The improved black kite optimization algorithm model randomly sets the black kite position as a potential solution in the search space to perform global search and local search, iteratively updates the black kite position to explore the optimal solution of the model parameters, and the improved black kite optimization algorithm model adopts a time-series adaptive attack strategy to dynamically adjust the attack step coefficient w. The expression for updating the position of a black kite individual is: in Indicates that the i-th black-winged kite updates the position of the j-th dimension in the t+1-th iteration, represents the position of the i-th black kite in the j-th dimension at the t-th iteration, t is the current iteration number, maxt is the maximum iteration number, p is a preset constant, r is a random number in [0, 1], and n represents the Cauchy mutation result.
4. The method for predicting operational management risks of small and scattered airlines based on fusion of multiple models according to claim 1 or 3 is characterized by: The improved black kite optimization algorithm model adopts a linear optimization strategy of equal dimension division. Each dimension d is evenly divided into n1 equal segments. The fitness value of each equal segment is calculated. By traversing the equal segment points on each dimension, the potential optimal solution on the dimension is explored, and the optimal equal segment point is selected to construct a new individual: Among them, ub(d) and lb(d) are the upper and lower bounds of the dimension, respectively, k∈[0,n1]; by changing the value of k, different equal division points can be obtained, and the individual fitness values corresponding to these equal division points can be calculated. The point with the best fitness value is selected as the representative of the dimension.
5. The method for predicting operational management risks of small and scattered airlines based on fusion of multiple models according to claim 3 is characterized by: The improved black kite optimization algorithm model introduces an improved differential evolution strategy, which designs a dynamic scaling factor F, which is expressed as follows: The improved black kite optimization algorithm model maintains a large mutation amplitude in the early stage of the algorithm to enhance the global search ability. As the iteration deepens, the mutation intensity is gradually reduced to conduct a fine search in the solution space. In the mutation operation, the elite individual is combined to generate the mutation vector v. i，g , the expression is: i，g =x r1，g +F·(x r2，g -x r3，g ), where x r1，g 、x r2，g 、x r3，g To randomly select three different individuals from the population; retain the dominant gene u through binomial crossover i，j，g , the expression is as follows: Where CR is the crossover probability, j rand is a randomly generated integer. This operation ensures that the mutation vector v i，j，g At least one dimension of information is retained, while retaining the original individual x i，j，g The new individual not only contains the information of the mutation, but also retains the dominant genes of the original individual.
6. The method for predicting operational management risks of small and scattered airlines based on fusion of multiple models according to claim 3 is characterized by: When the improved black kite optimization algorithm model detects that the fitness fluctuation is lower than the threshold δ, part of the population is reinitialized. The expression is as follows: x drs =x+r·(r·(ub+lb-x)-x), where x is the current individual, ub and lb are the upper and lower bounds of the individual value, and r is a random number in [0, 1]. By perturbing the current individual x, a new individual x is generated. drs ; Retain high-quality individuals through competitive selection, the expression is as follows: where f(x drs ) is used to generate a new individual x drs The fitness of f(x r ) is a randomly initialized individual x r The fitness of x r To randomly initialize individuals, by comparing x drs and x r fitness and select individuals with better fitness to ensure that high-quality individuals are always retained in the population.
7. The method for predicting operational management risks of small and scattered airlines based on fusion of multiple models according to claim 1 is characterized by: The safety risk indicator system includes four dimensional levels: event level, event type, incident stage and event cause. The event level includes several element levels such as general events and general event accident signs. The event type includes several element levels such as loss of control / stall, wind shear, bird strike, system failure / malfunction / jam. The incident stage includes several element levels such as taxiing, takeoff, climb, cruise and landing. The event cause includes several element levels such as air traffic control, maintenance, weather accidents and ground support.
8. The method for predicting operational management risks of small and scattered airlines based on fusion of multiple models according to claim 7 is characterized by: The safety risk index system includes four dimensional levels: event level, event type, incident stage and event cause, and the weights are determined by the Delphi method. The element levels under the four dimensional levels are also set with weights.
9. The method for predicting operational management risks of small and scattered airlines based on fusion of multiple models according to claim 3 is characterized by: The improved black kite optimization algorithm model selects a population leader to migrate the population. When the population fitness value under the population leader is lower than the threshold, the population leader will give up leadership and join the migrating population. The mutated population leader is obtained by using Cauchy mutation. Otherwise, the population leader will lead the population to migrate. The position expression of the mutated leader is as follows: where X leader For the leader before mutation, is the leader after mutation, γ is the set Cauchy distribution parameter, and Cauchy(0, 1) is a random number that obeys the standard Cauchy distribution.
10. The method for predicting operational management risks of small and scattered airlines based on fusion of multiple models according to claim 1 is characterized by: The time span group is divided by day, and the time span of the time span group is N1 days; or the time span group is divided by month, and the time span of the time span group is one natural month.