Method and system for predicting probability of occurrence of unsafe events of aircraft of research type

Abstract

The invention discloses a method and system for predicting the occurrence probability of unsafe events of a research-type aircraft, and the method comprises the steps: obtaining historical unsafe event data of the research-type aircraft before a research prediction day, and carrying out the collection of the historical unsafe event data according to a time period, calculating unsafe event probabilities of all time periods, and sequentially collecting the unsafe event probabilities according to a time sequence to form an unsafe event probability data set; constructing a combined prediction model comprising a bidirectional time convolution network BiTCN model, a bidirectional gating cycle unit and a LightGBM model, wherein the combined prediction model performs time series data prediction training by using an unsafe event probability data set; carrying out optimization processing on the combined prediction model by utilizing an improved kangaroo escape optimization algorithm model; and the combined prediction model outputs unsafe event probabilities of a plurality of time periods in the future according to the time sequence. According to the improved kangaroo escape optimization algorithm model, key hyper-parameter combinations in the combined prediction model can be efficiently searched, and the prediction precision and robustness of the combined prediction model are improved.

Description

Technical Field

[0001] This invention relates to the field of civil aviation aircraft incident risk prediction, and in particular to a method and system for predicting the probability of unsafe events occurring on a research aircraft model. Background Technology

[0002] Accurate prediction of the probability of unsafe events during civil aviation flights is crucial for ensuring flight safety. Unsafe events include flight accidents and incidents. Different aircraft models have different probabilities of unsafe events due to their performance characteristics. Even for the same aircraft model, the probability of unsafe events varies with weather and seasonal changes (e.g., strong winds and lightning in summer lead to unsafe events, while low temperatures and icy conditions in winter lead to unsafe events). In recent years, deep learning technology has been widely used in time series forecasting, providing new solutions for predicting the probability of unsafe events. Among them, recurrent neural networks (RNNs) and their variants (such as Long Short-Term Memory networks, LSTM) are widely used in wind shear and meteorological data forecasting due to their excellent performance in capturing nonlinear and temporal features. However, these network models rely heavily on the setting of hyperparameters, such as learning rate, number of hidden layer neurons, and batch size. Traditional manual parameter tuning methods are inefficient and difficult to find optimal solutions. In model learning and training, optimization algorithms can fine-tune hyperparameters. For example, the literature: Liu, S., et al. (2019). A Survey on DeepLearning with Attention Mechanism. IEEE Access, 7, 42023-42045, studies various hyperparameter optimization methods and emphasizes the importance of intelligent optimization algorithms in deep learning. There is an urgent need to research a method and system for predicting the probability of unsafe events, along with suitable model hyperparameter optimization methods and strategies, to achieve effective prediction of the probability of unsafe events. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for predicting the probability of unsafe events occurring on a research aircraft model. Based on historical unsafe event data, the probability of unsafe events is calculated by aggregating them according to time periods to obtain a time series unsafe event probability dataset. A combined prediction model is trained using the unsafe event probability dataset for time series data prediction. An improved kangaroo escape optimization algorithm model is used to optimize the model parameters of the combined prediction model, and finally, the optimal parameter combination that minimizes the prediction error is obtained. The optimized combined prediction model can accurately predict the probability of unsafe events in several consecutive time periods in the future.

[0004] The objective of this invention is achieved through the following technical solution: A method for predicting the probability of unsafe events occurring on a research aircraft model, the method comprising: S1. Obtain historical unsafe event data for the research aircraft type prior to the research prediction date and aggregate it by time period. Calculate the probability of unsafe events for each time period using the following formula: , The parameters are preset; the probabilities of unsafe events in each time period are sequentially collected in time series to form an unsafe event probability dataset; S2. Construct a combined prediction model that includes a bidirectional temporal convolutional network (BiTCN), a bidirectional gated recurrent unit (BiGRU), and a LightGBM model. The combined prediction model is trained using an insecure event probability dataset for time series data prediction. S3. Construct an improved kangaroo escape optimization algorithm model, and use the improved kangaroo escape optimization algorithm model to optimize the model parameters of the combined prediction model; the combined prediction model outputs the probability of unsafe events in several future time periods according to the time series.

[0005] To better implement this invention, the Bidirectional Temporal Convolutional Network (BiTCN) model captures contextual information and long-term dependency information of the insecure event probability dataset in both forward and backward directions through two parallel TCN layers (forward and backward); the Bidirectional Gated Recurrent Unit (BiGRU) further obtains the temporal contextual information of the insecure event probability dataset through two independent GRU layers (forward and backward); the LightGBM model uses gradient unilateral sampling to remove data with very small gradients and uses the remaining data to estimate information gain, employs an independent feature merging method to bundle mutually exclusive features to reduce the number of features, integrates multiple decision trees to accumulate the prediction results, and outputs the insecure event probability according to the time series prediction.

[0006] Preferably, in method S3, the improved kangaroo escape optimization algorithm model is used to optimize the model parameters of the combined prediction model according to the following method: S31. Combine the model parameters of the combined prediction model (BiTCN, BiGRU, and LightGBM) to form a hyperparameter combination, and use this hyperparameter combination as the initial position of kangaroo i in the population. The root mean square error between the predicted and actual values ​​is used as the fitness. ; S32. Determine the escape direction. The expression is as follows:

[0007] ,in For random vectors, It is the current best position in the global game. α represents the optimal local neighbor position of kangaroo i, where α, β, and γ are control parameters for adjusting randomness, global exploration, and local exploitation, respectively. The current global optimal fitness Let be the ratio of the fitness of kangaroo i to the current global best fitness; S33, Kangaroo i's location updated, new location. The expression is as follows:

[0008] ,in It is the step size factor. This represents the current iteration number. This represents the total number of iterations. Number of iterations The ratio of the fitness of kangaroo i to the current global best fitness. For the maximum ratio, The minimum ratio; when the number of iterations reaches the total number of iterations. If the fitness change is less than the set threshold, the optimization of the hyperparameter combination is stopped, and the hyperparameter combination at the global optimal position is output.

[0009] Preferably, fitness The expression is as follows: ,in For the true value, For predicted values, This represents the number of samples.

[0010] Preferably, the hyperparameter combination includes the hyperparameter combination of the Bidirectional Temporal Convolutional Network (BiTCN) model, the hyperparameter combination of the Bidirectional Gated Recurrent Unit (BiGRU) model, and the hyperparameter combination of the LightGBM model. The hyperparameter combination of the BiTCN model includes the number of layers, the kernel size, and the dilation factor. The hyperparameter combination of the BiGRU model includes the hidden layer size, the dropout rate, and the regularization parameter. The hyperparameter combination of the LightGBM model includes the maximum depth of the decision tree and the learning rate.

[0011] Preferably, the time period is divided into days, and the duration of the time period is N1 days; or the time period is divided into months, and the duration of the time period is one calendar month.

[0012] Preferably, in method S1, a probability dataset of unsafe events corresponding to each aircraft model is constructed according to the aircraft model under study. The probability dataset of unsafe events corresponding to each aircraft model is input into the combined prediction model and processed sequentially according to methods S2 to S3 to obtain the probability of unsafe events corresponding to each aircraft model under study. A chart and a comparison chart of the probability sequence data of unsafe events corresponding to all aircraft models under study are obtained through chart tools.

[0013] A prediction system for the probability of unsafe events occurring on a research aircraft type includes a data collection module, a data processing module, an improved kangaroo escape optimization algorithm model, and a combined prediction model. The data collection module is used to acquire historical unsafe event data of the research aircraft type before the research prediction date and aggregate it according to time periods. The data processing module calculates the probability of unsafe events for each time period according to the following formula: , The parameters are preset; the probabilities of unsafe events in each time period are sequentially collected in time series to form an unsafe event probability dataset; the combined prediction model includes a bidirectional temporal convolutional network (BiTCN) model, a bidirectional gated recurrent unit (BiGRU) model, and a LightGBM model. The combined prediction model is trained using the unsafe event probability dataset for time series data prediction; the improved kangaroo escape optimization algorithm model optimizes the model parameters of the combined prediction model to obtain the optimal hyperparameter combination. The combined prediction model uses the optimal hyperparameter combination to output the probabilities of unsafe events in several future time periods in time series.

[0014] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This invention calculates the probability of unsafe events by aggregating historical unsafe event data according to time periods to obtain a time series unsafe event probability dataset. The combined prediction model uses the unsafe event probability dataset for time series data prediction training. The bidirectional temporal convolutional network (BiTCN) model of the combined prediction model captures the contextual information and long-term dependency information of historical unsafe event data in both forward and backward directions through two parallel TCN layers (forward and backward). The bidirectional gated recurrent unit (BiGRU) of the combined prediction model uses two independent GRUs (forward and backward). The layer further acquires the time-series context information of historical unsafe event data; the LightGBM model of the combined prediction model uses gradient unilateral sampling to remove data with very small gradients and uses the remaining data to estimate information gain, and adopts the independent feature merging method to bundle mutually exclusive features to reduce the number of features. The LightGBM model integrates multiple decision trees to accumulate the prediction results and outputs them. The LightGBM model predicts the probability of unsafe events according to the time series; the improved kangaroo escape optimization algorithm is used to optimize the model parameters of the combined prediction model, and finally optimizes to obtain the optimal parameter combination that minimizes the prediction error. The optimized combined prediction model can accurately predict the probability of unsafe events in several consecutive time periods in the future.

[0015] (2) The present invention uses an improved kangaroo escape optimization algorithm model to optimize the model parameters of the combined prediction model, which can efficiently and automatically search for key hyperparameter combinations in the combined prediction model; in terms of convergence speed, it can find a better solution under the same number of iterations, thereby improving the prediction accuracy and robustness of the combined prediction model. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This example illustrates the time-series statistics of an insecurity event probability dataset obtained from historical insecurity event data. Figure 3 This is a schematic diagram comparing the prediction results with the actual results of the insecure event probability dataset divided into training set, validation set, and test set in the embodiment. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to embodiments: Example like Figure 1 As shown, a method for predicting the probability of unsafe events occurring on a research aircraft model is described, the method comprising: S1. Obtain historical unsafe event data for the research aircraft type prior to the research prediction date and aggregate it by time period. Calculate the probability of unsafe events for each time period using the following formula: , The parameters are preset. The probabilities of unsafe events in each time period are sequentially aggregated according to time series to form an unsafe event probability dataset. Preferably, the time period can be divided by days, and the duration of each time period is N1 days. Alternatively, the time period can be divided by months, and the duration of each time period is one calendar month. This embodiment takes dividing the time period by month (i.e., one time period is one calendar month) as an example, obtaining historical unsafe event data from January 2016 to September 2025 (example website is https: / / safety.caac.gov.cn / ), dividing it into 129 time periods by one calendar month, and calculating the probability of unsafe events for each of the 129 time periods. The calculation expression is as follows: , The example selects 1000, and calculates 129 months of unsafe event probability sequence data (i.e., unsafe event probability dataset). Time-series statistical sorting of the unsafe event probability sequence data yields the curve table shown below. Figure 2 As shown. Then, the insecurity event probability dataset is split into a training set (for model training and parameter optimization), a validation set, and a test set (for final evaluation).

[0018] S2. Construct a combined prediction model comprising a Bidirectional Temporal Convolutional Network (BiTCN), a Bidirectional Gated Recurrent Unit (BiGRU), and a LightGBM model. This combined prediction model is trained using an insecure event probability dataset for time-series data prediction. The BiTCN model of this invention captures contextual and long-term dependency information from both the forward and backward directions of the insecure event probability dataset through two parallel TCN layers (forward and backward). The TCN layers, by combining causal convolution and bidirectional information flow, achieve a more powerful feature extraction capability for time-series data. The architecture of the BiTCN model includes two parallel TCN layers (forward and backward), enabling it to simultaneously capture the "past" and "future" contextual information (i.e., contextual information from both the forward and backward directions) at each time point in the sequence. The TCN layers have dilated convolutional kernels and residual connection structures, possessing advantages such as parallel computation, stable gradients, and exponential growth of the receptive field, efficiently capturing long-term dependencies.

[0019] The Bidirectional Gated Recurrent Unit (BiGRU) further acquires time-series contextual information from the insecurity event probability dataset through two independent GRU layers: a forward layer and a backward layer. By integrating these two independent GRU layers, BiGRU comprehensively analyzes the time-series contextual information. The GRU layers themselves, through update and reset gate mechanisms, effectively control the flow and forgetting of information, mitigating the gradient vanishing problem. Compared to the LSTM model, it achieves similar performance with a simpler structure. BiGRU ensures that each output in the sequence depends not only on previous elements (forward layer) but also on subsequent elements (backward layer). Its bidirectional modeling capability allows it to obtain richer sequence representations than a unidirectional GRU, enabling it to perform judgment tasks based on complete context. The LightGBM model employs Gradient-based One-Side Sampling (GOSS) to remove data with very small gradients and uses the remaining data to estimate information gain (data with large gradients are more important for information gain, avoiding the influence of the long tail of low gradients). It uses Exclusive Feature Bundling (EFB) to bundle mutually exclusive features to reduce the number of features. The LightGBM model integrates multiple decision trees to accumulate the prediction results and outputs them. The LightGBM model predicts the probability of unsafe events according to time series.

[0020] S3. Construct an improved kangaroo escape optimization algorithm model and use the improved kangaroo escape optimization algorithm model to optimize the model parameters of the combined prediction model; the combined prediction model outputs the probability of unsafe events in several future time periods (the day of the research prediction and thereafter) according to the time series.

[0021] In some embodiments, the improved kangaroo escape optimization algorithm model IKEOA optimizes the model parameters of the combined prediction model according to the following method: S31. Combine the model parameters of the combined prediction model (BiTCN, BiGRU, and LightGBM) to form a hyperparameter combination, and use this hyperparameter combination as the initial position of kangaroo i in the population. In some embodiments, the hyperparameter combination includes the hyperparameter combination of the Bidirectional Temporal Convolutional Network (BiTCN) model, the hyperparameter combination of the Bidirectional Gated Recurrent Unit (BiGRU) model, and the hyperparameter combination of the LightGBM model. The hyperparameter combination of the BiTCN model includes the number of layers, kernel size, and dilation factor. The hyperparameter combination of the BiGRU model includes the hidden layer size, dropout rate, and regularization parameter. The hyperparameter combination of the LightGBM model includes the maximum depth of the decision tree and the learning rate.

[0022] This invention uses the root mean square error between the predicted and the true values ​​as the fitness. The preferred fitness of this invention The expression is as follows: ,in For the true value, For predicted values, The fitness function evaluates the quality of each set of parameters, guiding the improvement of the kangaroo escape optimization algorithm model to find the optimal parameter combination in the search space; through continuous iteration, the algorithm will gradually find the optimal combination of parameters that maximizes fitness. The minimum parameter combination is used to optimize the predictive performance of the combined prediction model.

[0023] S32. Determine the escape direction. The expression is as follows:

[0024] ,in For random vectors, It is the current best position in the global game. α represents the optimal local neighbor position of kangaroo i, where α, β, and γ are control parameters for adjusting randomness, global exploration, and local exploitation, respectively. The current global optimal fitness The fitness of kangaroo i is the ratio of its fitness to the current global best fitness. This invention improves the kangaroo escape optimization algorithm model IKEOA by using the ratio of kangaroo i's fitness to the current global best fitness. As the escape direction The automatic adjustment coefficient (specifically) and This enables intelligent direction switching, where poorer individuals tend to explore towards the global optimum, while better individuals focus more on local development; adaptive weights dynamically balance exploration and development, improving convergence efficiency.

[0025] S33, Kangaroo i's location updated, new location. The expression is as follows: , ,in It is the step size factor. This represents the current iteration number. This represents the total number of iterations. Number of iterations The ratio of the fitness of kangaroo i to the current global best fitness. Number of iterations The previous maximum ratio, Number of iterations The previous minimum ratio. This invention improves the kangaroo escape optimization algorithm model IKEOA during the position update phase, at the position of kangaroo i. Increase the number of iterations before The ratio of the fitness of kangaroo i to the current global best fitness. ,ratio Total number of iterations Previous maximum ratio Number of iterations Previous minimum ratio The algorithm is automatically adjusted so that the improved kangaroo escape optimization algorithm model IKEOA is given greater inertia in the early stage of iteration to enhance global exploration capabilities; and less inertia in the later stage of iteration to strengthen local development, effectively balancing the exploration and development of the algorithm.

[0026] When the number of iterations reaches the total number of iterations If the fitness change is less than a set threshold, the optimization process for the hyperparameter combination is stopped, and the hyperparameter combination at the globally optimal position is output as the model parameter combination for the combined prediction model. In the historical unsafe event data case from January 2016 to September 2025, the evaluation index of the combined prediction model for the test set in this embodiment is the root mean square error (RMSE). Figure 3 The RMSE was 0.016308, and the mean square error (MSE) was... Figure 3 The mean absolute error (MAE) is 0.000276. Figure 3 The MAE was 0.012913 and the coefficient of determination R² was... Figure 3 With an R² of 0.872573, plot a comparison chart of the predicted and actual values ​​on the test set (e.g., ...). Figure 3 As shown in the figure, the model performance is intuitively displayed. The improved kangaroo escape optimization algorithm model IKEOA of this invention can effectively transfer parameters from the initial random state and optimize them to the optimal solution, which significantly improves the prediction accuracy and robustness of the combined prediction model. The core of parameter transfer lies in the dynamic adjustment of the escape direction and position update formula, which ensures the balance between exploration and development of the improved kangaroo escape optimization algorithm model IKEOA.

[0027] In some embodiments, in method S1, unsafe event probability datasets corresponding to different aircraft models are constructed according to the aircraft model under study. These datasets are then input into a combined prediction model and processed sequentially according to methods S2 to S3 to obtain the unsafe event probabilities for each aircraft model. Finally, a graph of the unsafe event probability sequence data for all aircraft models is obtained using a graphing tool (the unsafe event probability sequence data curve for the same aircraft model is referenced). Figure 3 ) and comparative charts (comparing trends in the probability of unsafe events during flights of different research aircraft models through comparative analysis).

[0028] A prediction system for the probability of unsafe events occurring on a research aircraft model includes a data collection module, a data processing module, an improved kangaroo escape optimization algorithm model, and a combined prediction model. The data collection module acquires historical unsafe event data for the research aircraft model prior to the prediction date and aggregates it by time period. The data processing module calculates the probability of unsafe events for each time period using the following formula: , The parameters are preset. The probabilities of unsafe events for each time period are sequentially aggregated into an unsafe event probability dataset. The combined prediction model includes a bidirectional temporal convolutional network (BiTCN), a bidirectional gated recurrent unit (BiGRU), and a LightGBM model. This combined prediction model is trained using the unsafe event probability dataset for time-series data prediction. An improved kangaroo escape optimization algorithm is used to optimize the model parameters of the combined prediction model to obtain the optimal hyperparameter combination. The combined prediction model then uses this optimal hyperparameter combination to output the probabilities of unsafe events for several future time periods according to the time series.

[0029] 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 within the protection scope of the present invention.

Claims

1. A method for predicting the probability of unsafe events occurring on a research aircraft model, characterized in that: The methods include: S1. Obtain historical unsafe event data for the research aircraft type prior to the research prediction date and aggregate it by time period. Calculate the probability of unsafe events for each time period using the following formula: , The parameters are preset; the probabilities of unsafe events in each time period are sequentially collected in time series to form an unsafe event probability dataset; S2. Construct a combined prediction model that includes a bidirectional temporal convolutional network (BiTCN), a bidirectional gated recurrent unit (BiGRU), and a LightGBM model. The combined prediction model is trained using an insecure event probability dataset for time series data prediction. S3. Construct an improved kangaroo escape optimization algorithm model, and use the improved kangaroo escape optimization algorithm model to optimize the model parameters of the combined prediction model; the combined prediction model outputs the probability of unsafe events in several future time periods according to the time series.

2. The method for predicting the probability of unsafe events occurring on a research aircraft type according to claim 1, characterized in that: The Bi-directional Temporal Convolutional Network (BiTCN) model captures contextual information and long-term dependency information in both directions of the insecurity event probability dataset through two parallel TCN layers, forward and backward. The BiGRU (Bidirectional Gated Recurrent Unit) further acquires the time-series context information of the insecurity event probability dataset through two independent GRU layers, forward and backward. The LightGBM model uses gradient unilateral sampling to remove data with very small gradients and uses the remaining data to estimate information gain. It uses an independent feature merging method to bundle mutually exclusive features to reduce the number of features. The LightGBM model integrates multiple decision trees to accumulate the prediction results and outputs them. The LightGBM model predicts and outputs the insecurity event probability according to the time series.

3. The method for predicting the probability of unsafe events occurring on a research aircraft type according to claim 1, characterized in that: In method S3, the improved kangaroo escape optimization algorithm model is optimized by performing model parameter processing on the combined prediction model as follows: S31. Combine the model parameters of the combined prediction model (BiTCN, BiGRU, and LightGBM) to form a hyperparameter combination, and use this hyperparameter combination as the initial position of kangaroo i in the population. The root mean square error between the predicted and actual values ​​is used as the fitness. ; S32. Determine the escape direction. The expression is as follows: ； ,in For random vectors, It is the current best position in the global game. α represents the optimal local neighbor position of kangaroo i, where α, β, and γ are control parameters for adjusting randomness, global exploration, and local exploitation, respectively. The current global optimal fitness Let be the ratio of the fitness of kangaroo i to the current global best fitness; S33, Kangaroo i's location updated, new location. The expression is as follows: ； ,in It is the step size factor. This represents the current iteration number. This represents the total number of iterations. Number of iterations The ratio of the fitness of kangaroo i to the current global best fitness. For the maximum ratio, The minimum ratio; when the number of iterations reaches the total number of iterations. If the fitness change is less than the set threshold, the optimization of the hyperparameter combination is stopped, and the hyperparameter combination at the global optimal position is output.

4. The method for predicting the probability of unsafe events occurring on a research aircraft type according to claim 3, characterized in that: fitness The expression is as follows: ,in For the true value, For predicted values, This represents the number of samples.

5. The method for predicting the probability of unsafe events occurring on a research aircraft type according to claim 3, characterized in that: The hyperparameter combinations include the hyperparameter combinations of the Bidirectional Temporal Convolutional Network (BiTCN) model, the hyperparameter combinations of the Bidirectional Gated Recurrent Unit (BiGRU) model, and the hyperparameter combinations of the LightGBM model. The hyperparameter combinations of the BiTCN model include the number of layers, kernel size, and dilation factor. The hyperparameter combinations of the BiGRU model include the hidden layer size, dropout rate, and regularization parameter. The hyperparameter combinations of the LightGBM model include the maximum depth of the decision tree and the learning rate.

6. The method for predicting the probability of unsafe events occurring on a research aircraft type according to claim 1, characterized in that: The time period is divided into days, and the duration of the time period is N1 days; or the time period is divided into months, and the duration of the time period is one calendar month.

7. The method for predicting the probability of unsafe events occurring on a research aircraft type according to claim 1, characterized in that: In method S1, a probability dataset of unsafe events corresponding to each aircraft model under study is constructed. The probability dataset of unsafe events corresponding to each aircraft model under study is input into the combined prediction model and processed sequentially according to methods S2 to S3 to obtain the probability of unsafe events corresponding to each aircraft model under study. The probability sequence data charts and comparison charts of unsafe events corresponding to all aircraft models under study are obtained through charting tools.

8. A predictive system for the probability of unsafe events occurring on a research aircraft model, characterized in that: The system includes a data collection module, a data processing module, an improved kangaroo escape optimization algorithm model, and a combined prediction model. The data collection module is used to acquire historical unsafe event data of the research aircraft type before the research prediction date and aggregate it according to time periods. The data processing module calculates the probability of unsafe events for each time period using the following formula: , The parameters are preset; the probabilities of unsafe events in each time period are sequentially collected in time series to form an unsafe event probability dataset; the combined prediction model includes a bidirectional temporal convolutional network (BiTCN) model, a bidirectional gated recurrent unit (BiGRU) model, and a LightGBM model. The combined prediction model is trained using the unsafe event probability dataset for time series data prediction; the improved kangaroo escape optimization algorithm model optimizes the model parameters of the combined prediction model to obtain the optimal hyperparameter combination. The combined prediction model uses the optimal hyperparameter combination to output the probabilities of unsafe events in several future time periods in time series.

Images