Aging aircraft aviation safety prediction method based on improved WCGAN model

Abstract

The invention provides an aging aircraft aviation safety prediction method based on an improved WCGAN model, which utilizes small sample data in an open source database in an aviation safety voluntary report system to predict the aviation safety of an aging aircraft. Comprising the following steps: cleaning aircraft aviation fault event data in an open source database to obtain cleaned data; converting unstructured text report data in the cleaning data into structured data; a gradient penalty GP mechanism and a NeuralProphet algorithm are introduced to optimize the WCGAN model, and a WCGAN-GP model is obtained; parameters are optimized according to an error minimization principle; and inputting the small sample data into the well-trained WCGAN-GP model, and generating an aviation accident prediction value of the aging aircraft at a future moment. According to the method, the gradient penalty GP mechanism and the NeuralProphet algorithm are introduced to optimize and improve the WCGAN model, so that the defects of gradient disappearance, mode collapse, poor time sequence characteristic mining capability and the like in the training process of the traditional WCGAN model are overcome, the stability and the convergence speed of the improved WCGAN model are improved, and the improved WCGAN model is enabled to adapt to the characteristics of aviation accident data more effectively.

Description

Technical Field

[0001] The invention belongs to the technical field of aviation safety, and in particular relates to an aviation safety prediction method for aging aircraft based on an improved WCGAN model. Background Art

[0002] Accurate aviation accident prediction is of great significance to promoting proactive aviation safety management and improving industry efficiency, especially when aircraft enter the aging stage (referring to aircraft that have been in use for more than 75% of their life span). Aircraft body structure fatigue cracks, corrosion and other failures that affect aircraft safety occur frequently, seriously endangering flight safety.

[0003] Although the frequency of major aviation accidents has remained low in recent years with the innovation of aviation safety management concepts and technologies, according to public data from the International Civil Aviation Organization (ICAO), the number of fatal aviation accidents has not exceeded 6 per year since 2019. However, unlike general low-destructive and high-frequency aviation unsafe events, once an aviation accident occurs, it is extremely difficult to save it, and the loss of life and property is huge.

[0004] Therefore, preventing aviation accidents, especially preventing and controlling aviation accidents with serious consequences, is a major issue that concerns the long-term and overall situation, and highlights the urgency and importance of efficiently carrying out aviation accident prediction work.

[0005] After decades of development, the industry has mostly carried out statistical analysis of aviation safety incidents, and the development of aviation safety prediction theory has not yet fully matured, and there is still a lot of room for accurate prediction and intelligent decision-making in aviation safety. From the perspective of engineering practice and public literature, the current common prediction methods include time series prediction, econometric model prediction and machine learning prediction. In recent years, with the rise of cutting-edge technologies such as artificial intelligence and big data, machine learning prediction has been widely used. Deep learning has become an important branch of machine learning. It uses complex neural networks as its main architecture, and uses its powerful self-learning ability and nonlinear fitting ability to efficiently process complex data samples, so as to more fully explore the digital feature laws contained in the samples.

[0006] In deep learning, representative models include deep belief networks (DBN), recurrent neural networks (RNN), long short-term memory (LSTM), Transformer, etc. By combing and summarizing the relevant literature on aviation safety prediction, it is found that the objects of current aviation safety prediction research are mainly concentrated on samples of unsafe events or accident symptoms with rich data volume and significant digital features, and good prediction results are achieved through example verification. However, the objects of current aviation safety prediction research pay less attention to sparse and highly random aviation accident samples. This is because the mainstream prediction methods represented by deep learning have high requirements for data quality. Only by continuously optimizing model parameters from sufficient historical samples can a network structure with stable performance and strong generalization ability be obtained. However, small sample aviation accident data has the characteristics of uncertainty, complex causal relationships, and strong time series characteristics: (1) The characteristic of small sample aviation accident data is that the sample size is limited, resulting in high uncertainty in statistical analysis and model establishment. This uncertainty comes not only from insufficient data, but also from possible deviations in the data collection process; (2) The causes of aviation accidents are usually the result of the interaction of multiple factors, manifested as a complex causal network. Small sample data often find it difficult to reveal these complex relationships, resulting in traditional analysis methods being unable to identify key influencing factors; (3) Aviation accident data has significant time series characteristics, with data at different time points showing obvious dependencies and trends. The time series characteristics of small sample data are complex, requiring the prediction model to fully consider the impact of time factors on the occurrence of accidents. The above characteristics limit the computational efficiency of deep learning prediction models to a certain extent.

[0007] In the field of aviation safety prediction, few studies have attempted to build a deep learning model with good applicability. For some practical engineering problems in other aspects of prediction, considering the high uncertainty, poor generalization ability and weak interpretability of the directly constructed analytical mapping model in existing small sample prediction research, some scholars have introduced a new generative adversarial network (GAN) deep learning model. The GAN deep learning model generates realistic data in unsupervised learning through adversarial training of the generator and the discriminator, thereby expanding the small sample data set and improving the generalization ability of the GAN deep learning model. In recent years, the GAN deep learning model has achieved good results for small sample prediction tasks.

[0008] How to design a small sample prediction model to predict aviation safety of aging aircraft based on the actual characteristics of aviation accident data has become an urgent problem to be solved. Summary of the invention

[0009] The purpose of the present invention is to solve the shortcomings of the prior art that aircraft aviation safety prediction cannot be performed based on small samples, and to provide an aviation safety prediction method for aging aircraft based on an improved WCGAN model, introduce a gradient penalty GP mechanism and a NeuralProphet algorithm to optimize the improved WCGAN model, overcome the defects of the traditional WCGAN model such as gradient disappearance, mode collapse and lack of time series feature mining ability during training, improve the stability and convergence speed of the improved WCGAN model, and make the improved WCGAN model more effectively adapt to the characteristics of aviation accident data.

[0010] To achieve the above purpose, the technical solution provided by the present invention is:

[0011] A method for predicting aviation safety of aging aircraft based on an improved WCGAN model includes: using small sample data from an open source database in an aviation safety voluntary reporting system to predict aviation safety of aging aircraft; specifically including:

[0012] Data processing: clean the aircraft aviation failure event data in the open source database to obtain clean data; convert the unstructured text report data in the clean data into structured data;

[0013] Improved WCGAN model: The gradient penalty GP mechanism and NeuralProphet algorithm were introduced to optimize the WCGAN model and obtain the WCGAN-GP model;

[0014] Training the WCGAN-GP model: The WCGAN-GP model traverses the data and optimizes the parameters based on the principle of minimizing the error;

[0015] Testing the WCGAN-GP model: inputting the small sample data into the maturely trained WCGAN-GP model, the WCGAN-GP model generates aviation accident prediction values ​​for aging aircraft at future times.

[0016] As a further limitation of the present invention, the data processing specifically includes:

[0017] Using open source data in the aviation safety voluntary reporting system, screen and compile statistics on aviation accident data required for research. The aviation accident data is cleaned data.

[0018] The time range of the aviation accident data required for the study is further screened and statistics are made on a monthly basis, thereby converting the unstructured text report data into structured text report data.

[0019] As a further limitation of the present invention, the improved WCGAN model specifically includes:

[0020] The WCGAN-GP model includes a generator and a discriminator. The cleaned data includes historical aviation accident record data and causal event data affecting aviation accidents. The historical aviation accident record data is represented as real data x r , the causal event data affecting aviation accidents is represented as conditional input y; random noise z is concatenated with the conditional input y and input into the generator, and the generator outputs an aviation accident prediction value G(z|y); the aviation accident prediction value G(z|y) is compared with the real data x r Input the discriminator, the discriminator not only determines the aviation accident prediction value G(z|y) and the real data x r The discriminator further determines whether the aviation accident prediction value G(z|y) satisfies the constraint of the conditional input y; specifically:

[0021] The game objective function with conditional probability distribution is expressed as:

[0022]

[0023] In the formula, Represents the real data sample p r The expected value of (x), Represents the generated data sample The expected value of , D(.) represents the output of the discriminator;

[0024] The Wasserstein distance is expressed as:

[0025]

[0026] In the formula, express With p r The set of all possible joint distributions of (x), ∫ x×y d(x,y)dγ(x,y) represents the expected value of the sample distance, and its lower bound is obtained from the expected value With p r (x) is the Wasserstein distance between them.

[0027] As a further limitation of the present invention, the gradient penalty GP mechanism is introduced into the loss function of the discriminator to constrain the gradient of the discriminator to a controllable range; the gradient penalty GP mechanism calculates the output value of the discriminator for the interpolated data Gradient The square of the deviation from the unit length makes the gradient value approach the unit value while ensuring that the penalty term is always positive, and the greater the deviation, the greater the penalty; where:

[0028] The GP expression of the gradient penalty GP mechanism is:

[0029]

[0030] In the formula, λ represents the adjustment parameter for controlling the intensity of the penalty. represents the linear interpolation between the generated samples and the real samples, represents the output value of the discriminator for the interpolated data, ||.||2 represents the L2 norm, ε represents a random number that obeys a uniform distribution, and x represents the real data;

[0031] The gradient penalty GP mechanism is introduced, and the objective function expression in the WCGAN-GP model is:

[0032]

[0033] As a further limitation of the present invention, the expression of the NeuralProphet algorithm is:

[0034]

[0035] In the formula, T(t) represents the trend value, S(t) represents the seasonal effect, E(t) represents the holiday effect, F(t) represents the regression effect of future known exogenous variables, A(t) represents the autoregressive effect of historical observed variables, and L(t) represents the regression effect of lagged observations of exogenous variables;

[0036] Among them: 6 components are configured individually or in combination to form a NeuralProphet parsing model, which contains multiple neural network hidden layers.

[0037] Assume that there is time series data X = [x1, x2, ..., x n ], input the hidden layer of the neural network, expressed as:

[0038]

[0039] In the formula, H represents the output result of the time series data passing through the hidden layer of the neural network, ω i represents the weight, b i represents the bias term, f(.) represents the activation function;

[0040] The output result is passed through a linear output layer to predict the aviation accident prediction value at the future time, which is expressed as:

[0041]

[0042] Where Y represents the output of the linear layer, ω j represents the weight, H j represents the implicit output result layer, b j Represents the bias term.

[0043] As a further limitation of the present invention, the training of the WCGAN-GP model specifically includes:

[0044] The number of hidden layers of the network model is traversed in a step search form, and the Adam algorithm is used to adaptively adjust the learning rate of the neural network. The dropout regularization mechanism is added between the hidden layers of the discriminator to reduce the redundancy of the model.

[0045] Compare the mean square errors under different parameter combinations and optimize the parameters based on the principle of minimizing the error.

[0046] As a further limitation of the present invention, it also includes:

[0047] Evaluation and prediction: Based on the aviation accident prediction value, the feasibility and effectiveness of aviation safety prediction for aging aircraft are evaluated; among which:

[0048] The mean square root error, mean absolute error, and mean absolute percentage error are used as evaluation indicators to evaluate the prediction ability of the WCGAN-GP model, which is expressed as:

[0049]

[0050] In the formula, RMSE represents the mean square root error index, y i represents the i-th sample in the validation data set, y′ i Represents the aviation accident prediction value, n represents the number of samples, MAE represents the mean absolute error index, and MAPE represents the mean absolute percentage index.

[0051] The advantages of the present invention are:

[0052] 1. The present invention introduces the gradient penalty GP mechanism and the NeuralProphet algorithm to optimize and improve the WCGAN model, overcomes the defects of the traditional WCGAN model such as gradient vanishing, mode collapse and lack of time series feature mining ability during training, improves the stability and convergence speed of the improved WCGAN model, and makes the improved WCGAN model more effectively adapt to the characteristics of aviation accident data.

[0053] 2. The present invention utilizes the open source data in the aviation safety voluntary reporting system, highlights the mining of the time series characteristics of accident data samples in the open source data, and uses the WCGAN-GP model structure optimization and NeuralProphet algorithm to enhance the recognition ability of autocorrelation, trend and periodicity in the time series distribution of time series characteristics. The prediction results are more accurate and reliable, effectively improving the prediction accuracy and robustness of aviation accident data.

[0054] 3. The present invention has a higher prediction effect on small sample aviation accident data, can predict aviation accidents more accurately, and has better performance in solving the uncertainty of small sample data, complex causal relationships and strong time series characteristics.

[0055] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0057] Figure 1 :The present invention provides a method for predicting aviation safety of aging aircraft based on an improved WCGAN model Figure 1 ;

[0058] Figure 2 : The sample example table of extracting safety report from ASRS provided by the present invention;

[0059] Figure 3 : The monthly statistical results of aviation safety data collected by the present invention are shown in figure;

[0060] Figure 4 : The main parameter value table of the NeuralProphet algorithm provided by the present invention;

[0061] Figure 5 : The equipment failure event prediction result diagram provided by the present invention;

[0062] Figure 6 : The program inappropriate event prediction result diagram provided by the present invention;

[0063] Figure 7 : The miscommunication event prediction result diagram provided by the present invention;

[0064] Figure 8 : The situational awareness deviation event prediction result diagram provided by the present invention;

[0065] Fig. 9 :The present invention provides a method for predicting aviation safety of aging aircraft based on an improved WCGAN model Figure 2 ;

[0066] Fig.10 : A comparison result diagram of the small sample aviation accident prediction method provided by the present invention;

[0067] Fig.11 : Comparison table of RMSE, MAPE and MAE of different methods provided by the present invention;

[0068] Fig.12 : Illustration of the generator and discriminator structures in the improved WCGAN model provided by the present invention. DETAILED DESCRIPTION

[0069] Embodiments of the present invention are described in detail below. The embodiments are exemplary and intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0070] The embodiment of the present invention discloses a method for predicting aviation safety of aging aircraft based on an improved WCGAN model, including: using small sample data in an open source database in an aviation safety voluntary reporting system to predict aviation safety of aging aircraft; Figure 1 The aviation safety prediction method for aging aircraft according to the embodiment of the present invention specifically includes: (1) data processing, (2) improving the WCGAN model, (3) training the WCGAN-GP model, and (4) testing the WCGAN-GP model.

[0071] in:

[0072] (1) Data processing specifically includes: cleaning the aircraft aviation failure event data in the open source database to obtain cleaned data; and converting the unstructured text report data in the cleaned data into structured data.

[0073] More specifically, the above data processing includes: using the open source data in the Aviation Safety Reporting System (ASRS), filtering and counting the aviation accident data required for the research according to the "events" attribute value, and the aviation accident data is cleaned data; further filtering the time range of the aviation accident data required for the research according to the "date" attribute value, and counting it on a monthly basis, thereby converting the unstructured text report data into structured text report data.

[0074] More specifically, the research data of the embodiment of the present invention is selected from the voluntary report text records recorded by the Aviation Safety Voluntary Reporting System (ASRS for short), including two types of samples: accident data records that cause functional damage to aircraft (aviation accidents for short) and unsafe incident data records. The number of occurrences is counted by month based on the "time" attribute in the original text record, thereby structuring the text data, with the number of aviation accidents serving as the output indicator. Unsafe incidents are risk sources that trigger accidents. Although they do not directly cause established losses, they have a causal relationship with accidents. The embodiment of the present invention uses the number of unsafe incidents as an input indicator. Please refer to Figure 2 , Figure 2 The Sample Safety Report Extracted from ASRS table shows some examples of attribute information for a voluntary report sample.

[0075] The embodiment of the present invention can extract the monthly statistics of aviation accidents from July 2017 to June 2021 from the ASRS database, such as Figure 3 shown.

[0076] (2) The improvement of the WCGAN model is as follows: the gradient penalty GP mechanism and the NeuralProphet algorithm are introduced to optimize the WCGAN model to obtain the WCGAN-GP model.

[0077] The WCGAN model of the embodiment of the present invention adds a conditional input y to the original GAN ​​structure, so that the generator can generate samples of a specific category according to given conditions, thereby enhancing the flexibility and practicality of the generator. The wasserstein distance is used instead of the JS distance in the traditional GAN ​​as the loss function for model training. This distance can more accurately measure the difference between the generated samples and the real samples, thereby improving the performance of the generator and the stability of training. Specifically, it includes:

[0078] See also Fig.12 The generator of the embodiment of the present invention includes an input layer, a feedforward neural network 1, a feedforward neural network 2, ... a feedforward neural network n, and a fully connected layer. The discriminator of the embodiment of the present invention includes an input layer, a concatenation layer, a feedforward neural network 1, a feedforward neural network 2, ... a feedforward neural network n, and a fully connected layer.

[0079] The embodiment of the present invention uses a step search to traverse the number of hidden layers of the network model, uses the Adam algorithm to adaptively adjust the network learning rate, and adds a dropout regularization mechanism between the hidden layers of the neural network of the discriminator to reduce the redundancy of the model. Compare the mean square error (MSE) under different parameter combinations, and optimize the parameters based on the principle of minimizing the error. The mean square error expression is:

[0080]

[0081] In the formula, y i represents the i-th sample in the validation data set, y′ i represents the predicted value of aviation accidents, and n represents the number of samples.

[0082] In the small sample aviation accident prediction task, the data set is extracted from the historical aviation accidents of old aircraft and their related influencing factors. Among them: the WCGAN-GP model includes a generator and a discriminator, and the cleaned data includes historical aviation accident record data and the data of the inducing events that affect aviation accidents. The historical aviation accident record data is represented as the real data x r, the data of the inducing events that affect aviation accidents are represented as conditional input y; random noise z is concatenated with the conditional input y and input into the generator, and the generator outputs the aviation accident prediction value G(z|y); the aviation accident prediction value G(z|y) is compared with the real data x r Input discriminator, the discriminator not only judges the aviation accident prediction value G(z|y) and the real data x r The discriminator also determines whether the aviation accident prediction value G(z|y) satisfies the constraints of the conditional input y; specifically:

[0083] The game objective function with conditional probability distribution is expressed as:

[0084]

[0085] In the formula, Represents the real data sample p r The expected value of (x), Represents the generated data sample The expected value of , D(.) represents the output of the discriminator;

[0086] When measuring the distribution similarity between generated samples and real samples, traditional GAN ​​models usually use JS divergence as a metric, as shown in the following formula:

[0087]

[0088] Combining the above two equations, the objective function of network training can be rewritten as:

[0089]

[0090] If the generated sample distribution does not overlap with the real sample distribution at all, the JS divergence will be a constant, resulting in a gradient of 0 and failure to update normally. In view of the defects of JS divergence, the embodiment of the present invention uses Wasserstein distance as a metric. The basic idea of ​​Wasserstein distance is to generate sample distribution Move to the true sample distribution p r (x) The minimum cost required to measure the similarity between two distributions.

[0091] The Wasserstein distance is expressed as:

[0092]

[0093] In the formula, express With p r The set of all possible joint distributions of (x), ∫ x×yd(x,y)dγ(x,y) represents the expected value of the sample distance, and its lower bound is obtained from the expected value With p r (x) is the Wasserstein distance between them.

[0094] Furthermore, the above-mentioned gradient penalty GP mechanism of the embodiment of the present invention is introduced into the loss function of the discriminator, so that the gradient of the discriminator is constrained to a controllable range; the gradient penalty GP mechanism calculates the output value of the discriminator for the interpolation data Gradient The square of the deviation from the unit length makes the gradient value approach the unit value while ensuring that the penalty term is always positive, and the greater the deviation, the greater the penalty; where:

[0095] The GP expression of the gradient penalty GP mechanism is:

[0096]

[0097] In the formula, λ represents the adjustment parameter for controlling the intensity of the penalty. represents the linear interpolation between the generated samples and the real samples, represents the output value of the discriminator for the interpolated data, ||.||2 represents the L2 norm, ε represents a random number that obeys a uniform distribution, and x represents the real data;

[0098] The gradient penalty GP mechanism is introduced, and the objective function expression in the WCGAN-GP model is:

[0099]

[0100] Based on the WCGAN model, the NeuralProphet selected in the embodiment of the present invention is compatible with the excellent nonlinear fitting ability of the neural network and the characteristics of Prophet's efficient learning of time series patterns, and is suitable for complex time series prediction tasks. Using the NeuralProphet dependency package, the training set in the statistical aviation unsafe data is used to optimize the model structure, and then the corresponding NeuralProphet time series prediction model is established for each input indicator. The optimization method combining step search and empirical assignment is adopted to finally determine the main parameter value table of the NeuralProphet algorithm as shown in the following figure. Figure 4 As shown. Figure 4 The parameter settings in the table record the changes in the loss of the NeuralProphet model corresponding to the four input indicators in each training cycle.

[0101] The expression of the NeuralProphet algorithm in the embodiment of the present invention is:

[0102]

[0103] In the formula, T(t) represents the trend value, S(t) represents the seasonal effect, E(t) represents the holiday effect, F(t) represents the regression effect of future known exogenous variables, A(t) represents the autoregressive effect of historical observed variables, and L(t) represents the regression effect of lagged observations of exogenous variables;

[0104] Among them: 6 components are configured individually or in combination to form the NeuralProphet analysis model. Under the default setting, only the trend component and seasonal effect are used. The NeuralProphet analysis model contains multiple neural network hidden layers.

[0105] Assume that there is time series data X = [x1, x2, ..., x n ], input to the hidden layer of the neural network, expressed as:

[0106]

[0107] In the formula, H represents the output result of the time series data passing through the hidden layer of the neural network, ω i represents the weight, b i represents the bias term, f(.) represents the activation function;

[0108] The output result is passed through a linear output layer to predict the aviation accident prediction value at the future time, which is expressed as:

[0109]

[0110] Where Y represents the output of the linear layer, ω j represents the weight, H j represents the implicit output result layer, b j Represents the bias term.

[0111] (3) Training the WCGAN-GP model is as follows: The WCGAN-GP model traverses the data and optimizes the parameters based on the principle of minimizing the error.

[0112] Specifically include:

[0113] Firstly, the number of hidden layers of the network model is traversed in the form of step search, and the Adam algorithm is used to adaptively adjust the network learning rate. Then, the dropout regularization mechanism is added between the hidden layers of the discriminator to reduce the redundancy of the model. Finally, the mean square error (MSE) under different parameter combinations is compared, and the parameters are optimized based on the principle of minimizing the error.

[0114] (4) Testing the WCGAN-GP model involves inputting small sample data into the well-trained WCGAN-GP model, which generates aviation accident prediction values ​​for aging aircraft at future times.

[0115] Test results such as Fig.10 As shown, from Fig.10 It can be seen that: (1) Compared with the traditional WCGAN model, the improved WCGAN-GP model proposed in this invention has significantly improved the prediction performance on the test set, and both the absolute value and the peak trend are more consistent with the test sample (Test). (2) The neural network models (NeuralProphet, LSTM and TCN) of the time series prediction class and the decision tree models (RF, XGboost and LightGBM) of the ensemble learning class have large prediction errors on most test samples and are insufficient in describing the changing trend. This indirectly shows that the above comparison methods have the shortcomings of insufficient learning of time series rules and inaccurate description of causal mapping relationships in the problem of small sample time series prediction.

[0116] The embodiment of the present invention calculates the predicted values ​​of the input indicators at four test time points under the well-trained NeuralProphet model. The NeuralProphet model enhances the ability to recognize time series characteristics by combining the deep learning architecture with the traditional time series prediction method. Its mechanism includes automatic modeling of seasonality and trends, using time feature embedding, adopting adaptive learning rate, handling missing values ​​and outliers, considering the interaction effects of multiple features, and providing interpretability. These features enable NeuralProphet to more accurately capture complex time series relationships during the prediction process, thereby improving the accuracy of the prediction results. The calculation results are as follows: Figure 5-Figure 8 shown.

[0117] See also Fig. 9 The aging aircraft aviation safety prediction method of the embodiment of the present invention includes: (1) data processing, (2) improving the WCGAN model, (3) training the WCGAN-GP model, and (4) testing the WCGAN-GP model, and also includes: (5) evaluation prediction.

[0118] (5) The evaluation and prediction are as follows: Based on the aviation accident prediction value, the feasibility and effectiveness of the aviation safety prediction of aging aircraft are evaluated; wherein: the mean square root error, mean absolute error, and mean absolute percentage error are used as evaluation indicators to evaluate the prediction ability of the WCGAN-GP model, which is expressed as:

[0119]

[0120] In the formula, RMSE represents the mean square root error index, y irepresents the i-th sample in the validation dataset, y′ i Represents the aviation accident prediction value, n represents the number of samples, MAE represents the mean absolute error index, and MAPE represents the mean absolute percentage index.

[0121] In order to more intuitively evaluate the prediction effect of the method disclosed in the present invention, the prediction methods of WCGAN, NeuralProphet, LSTM, TCN, RF, XGBoost and LightGBM are used as control models of the prediction method of the present invention, and the prediction results are calculated in the same experimental environment. Fig.10 shown.

[0122] from Fig.10 It can be seen that (1) compared with the traditional WCGAN model, the improved WCGAN-GP model disclosed in the present invention has significantly improved the prediction performance on the test set, and both the absolute value and the peak trend are more consistent with the test sample (Test). (2) The neural network models (NeuralProphet, LSTM and TCN) of the time series prediction class and the decision tree models (RF, XGboost and LightGBM) of the ensemble learning class have large prediction errors on most test samples and are insufficient in describing the changing trends. This indirectly shows that the above comparison methods have the shortcomings of insufficient learning of time series rules and inaccurate description of causal mapping relationships in the problem of small sample time series prediction.

[0123] To further quantitatively evaluate the generalization performance of the model, record the RMSE, MAE, and MAPE of the model in each test sample. Fig.11 ,right Fig.11 Comparative analysis shows that in the performance of the prediction model, the improved WCGAN-GP disclosed in the embodiment of the present invention always maintains the best accuracy, and performs well in RMSE, MAE and MAPE indicators, which is more than 30% higher than the unimproved WCGAN model. In addition, the decision tree model of the ensemble learning class is generally better than the neural network model of the time series prediction class in terms of prediction accuracy.

[0124] Through the analysis of the prediction model operation results, it can be seen that the improved WCGAN-GP model disclosed in the embodiment of the present invention can effectively predict the changing trend of small sample aviation accidents, warn of the occurrence of major dangerous consequences in advance, and help managers take targeted protective measures in a timely manner.

[0125] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should be included in the protection scope of the present invention.

Claims

1. A method for predicting aviation safety of aging aircraft based on an improved WCGAN model, characterized in that: include: Use small sample data from the open source database of the Aviation Safety Voluntary Reporting System to predict aviation safety for aging aircraft; Specifically include: Data processing: clean the aircraft aviation failure event data in the open source database to obtain clean data; convert the unstructured text report data in the clean data into structured data; Improved WCGAN model: The gradient penalty GP mechanism and NeuralProphet algorithm were introduced to optimize the WCGAN model and obtain the WCGAN-GP model; Training the WCGAN-GP model: The WCGAN-GP model traverses the data and optimizes the parameters based on the principle of minimizing the error; Testing the WCGAN-GP model: inputting the small sample data into the maturely trained WCGAN-GP model, the WCGAN-GP model generates aviation accident prediction values ​​for aging aircraft at future times.

2. The method for predicting aviation safety of aging aircraft based on the improved WCGAN model according to claim 1, characterized in that: The data processing specifically includes: Using open source data in the aviation safety voluntary reporting system, screen and compile statistics on aviation accident data required for research. The aviation accident data is cleaned data. The time range of the aviation accident data required for the study is further screened and statistics are made on a monthly basis, thereby converting the unstructured text report data into structured text report data.

3. The method for predicting aviation safety of aging aircraft based on the improved WCGAN model according to claim 1, characterized in that: The improved WCGAN model specifically includes: The WCGAN-GP model includes a generator and a discriminator. The cleaned data includes historical aviation accident record data and causal event data affecting aviation accidents. The historical aviation accident record data is represented as real data x r , the causal event data affecting aviation accidents is represented as conditional input y; random noise z is concatenated with the conditional input y and input into the generator, and the generator outputs an aviation accident prediction value G(z|y); the aviation accident prediction value G(z|y) is compared with the real data x r Input the discriminator, the discriminator not only determines the aviation accident prediction value G(z|y) and the real data x r The discriminator also determines whether the aviation accident prediction value G(z|y) satisfies the constraint of the conditional input y; specifically: The game objective function with conditional probability distribution is expressed as: In the formula, Represents the real data sample p r The expected value of (x), Represents the generated data sample The expected value of , D(.) represents the output of the discriminator; The Wasserstein distance is expressed as: In the formula, express With p r The set of all possible joint distributions of (x), ∫ x×y d(x,y)dγ(x,y) represents the expected value of the sample distance, and its lower bound is obtained from the expected value With p r (x) is the Wasserstein distance between them.

4. The method for predicting aviation safety of aging aircraft based on the improved WCGAN model according to claim 3 is characterized in that: The gradient penalty GP mechanism is introduced into the loss function of the discriminator to constrain the gradient of the discriminator to a controllable range; the gradient penalty GP mechanism calculates the output value of the discriminator for the interpolated data Gradient The square of the deviation from the unit length makes the gradient value approach the unit value while ensuring that the penalty term is always positive, and the greater the deviation, the greater the penalty; where: The GP expression of the gradient penalty GP mechanism is: In the formula, λ represents the adjustment parameter that controls the intensity of the penalty. represents the linear interpolation between the generated samples and the real samples, represents the output value of the discriminator for the interpolated data, ||.||2 represents the L2 norm, ε represents a random number that obeys a uniform distribution, and x represents the real data; The gradient penalty GP mechanism is introduced, and the objective function expression in the WCGAN-GP model is:

5. The method for predicting aviation safety of aging aircraft based on the improved WCGAN model according to claim 1, characterized in that: The expression of NeuralProphet algorithm is: In the formula, T(t) represents the trend value, S(t) represents the seasonal effect, E(t) represents the holiday effect, F(t) represents the regression effect of future known exogenous variables, A(t) represents the autoregressive effect of historical observed variables, and L(t) represents the regression effect of lagged observations of exogenous variables; Among them: 6 components are configured individually or in combination to form the NeuralProphet analytical model. The NeuralProphet parsing model contains multiple neural network hidden layers. Assume that there is time series data X = [x1, x2, ..., x n ], input the hidden layer of the neural network, expressed as: In the formula, H represents the output result of the time series data passing through the hidden layer of the neural network, ω i represents the weight, b i represents the bias term, f(.) represents the activation function; The output result is passed through a linear output layer to predict the aviation accident prediction value at the future time, which is expressed as: Where Y represents the output of the linear layer, ω j represents the weight, H j represents the implicit output result layer, b j Represents the bias term.

6. The method for predicting aviation safety of aging aircraft based on the improved WCGAN model according to claim 1, characterized in that: The training of the WCGAN-GP model specifically includes: The number of hidden layers of the network model is traversed in a step search form, and the Adam algorithm is used to adaptively adjust the learning rate of the neural network. The dropout regularization mechanism is added between the hidden layers of the discriminator to reduce the redundancy of the model. Compare the mean square errors under different parameter combinations and optimize the parameters based on the principle of minimizing the error.

7. The method for predicting aviation safety of aging aircraft based on the improved WCGAN model according to claim 1, characterized in that: Also includes: Evaluation and prediction: Based on the aviation accident prediction value, the feasibility and effectiveness of aviation safety prediction for aging aircraft are evaluated; among which: The mean square root error, mean absolute error, and mean absolute percentage error are used as evaluation indicators to evaluate the prediction ability of the WCGAN-GP model, which is expressed as: In the formula, RMSE represents the mean square root error index, y i Represents the i-th sample in the validation data set, y' i Represents the aviation accident prediction value, n represents the number of samples, MAE represents the mean absolute error index, and MAPE represents the mean absolute percentage index.

Images