Method and device for identifying key events of a high-speed vehicle based on a sequence of infrared radiations

Abstract

The application discloses a method and device for identifying key events of a high-speed aircraft based on an infrared radiation sequence, which can realize accurate identification of key events of the high-speed aircraft under limited data conditions. The method comprises the following steps: (1) establishing a mass center motion equation of the high-speed aircraft to obtain a flight trajectory of the high-speed aircraft in a boost phase; (2) constructing an infrared radiation model of a tail flame to generate an infrared radiation time sequence; (3) superimposing a Gaussian random number and a triangular wave on the infrared radiation sequence; (4) introducing an augmentation method of the infrared radiation sequence data on the basis of a basic sample; (5) identifying events by using an XGBoost regression model, capturing long-time dependent relationships in the infrared radiation sequence by using a multi-layer expansion convolution structure of a TCN, introducing a Dropout mechanism in a convolution layer, and adopting an Adam optimization algorithm for model parameter optimization; (6) stacking the basic model results of different models as input, and adopting a multi-layer perception to perform stacking; and (7) error evaluation.

Description

Technical Field

[0001] This invention relates to the technical field of target infrared radiation simulation and target event recognition, and in particular to a method for recognizing critical events of high-speed aircraft based on infrared radiation sequences, and a device for recognizing critical events of high-speed aircraft based on infrared radiation sequences. Background Technology

[0002] In infrared detection systems, high-speed aircraft generate significant infrared radiation signals from their exhaust plumes during the boost phase. These signals are typically used for early target detection and tracking. Infrared radiation time series not only reflect changes in the aircraft's motion state but also contain crucial information about the propulsion system's operation, such as engine shutdown events and deflagration separation events during flight.

[0003] However, in actual infrared observation, real infrared observation data is difficult to obtain due to high experimental costs and strong data confidentiality requirements. Furthermore, infrared signals are typically affected by various factors such as background radiation, atmospheric transmission, and sensor noise, increasing the difficulty of identifying critical events. Traditional methods based on threshold detection or empirical features struggle to accurately identify critical events in complex infrared time series. Existing research also largely focuses on target detection, with relatively little research on critical event identification during high-speed aircraft flight.

[0004] In recent years, machine learning methods have demonstrated excellent performance in modeling complex nonlinear data. However, machine learning models typically rely on large amounts of training data, while infrared observation data is difficult to acquire. Therefore, how to accurately identify key events of high-speed aircraft under limited data conditions has become an important research problem in the field of infrared detection. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the technical problem to be solved by the present invention is to provide a method for identifying key events of high-speed aircraft based on infrared radiation sequences, which can achieve accurate identification of key events of high-speed aircraft under limited data conditions.

[0006] The technical solution of this invention is: a method for identifying critical events in high-speed aircraft based on infrared radiation sequences, comprising the following steps: (1) Establish the motion equation of the center of mass of the high-speed aircraft, describe the motion state of the aircraft through the forces such as thrust, aerodynamics and gravity, and solve the motion equation by numerical integration method to obtain the flight trajectory of the aircraft during the boost phase. (2) Based on flight trajectory data, construct an infrared radiation model of the tail flame, calculate the infrared radiation intensity through tail flame geometry, temperature distribution and observation distance, and generate an infrared radiation time series; (3) A Gaussian random number with a mean of zero is superimposed on each data point of the infrared radiation sequence, and its standard deviation is adjusted by the signal amplitude to simulate the detection under different signal-to-noise ratio conditions; a periodic triangular wave is added to the signal, which is an asymmetric triangular wave, and perturbation is added to simulate periodic interference. (4) Based on the basic sample, an augmentation method for infrared radiation sequence data is introduced to construct a time mapping function. A mapping from the original time t to the new time t' is established using cubic spline interpolation. (5) The XGBoost regression model is used to identify events. To adapt to the input format of the XGBoost model, a sliding window is used to extract features and flatten the time series. To improve the generalization ability of the model and prevent overfitting, the key parameters of the XGBoost model are set. To fully learn the features of the infrared radiation sequence, the multi-layer dilated convolution structure of TCN is used to gradually increase the receptive field by increasing the dilation coefficient and capture the long-term dependence in the infrared radiation sequence. During the network training process, in order to improve the generalization ability of the model, the Dropout mechanism is introduced in the convolutional layer with a dropout rate of 0.2. The Adam optimization algorithm is used to optimize the model parameters with a learning rate of 0.001. (6) Stacking takes the results of the base models of different models as input and uses a multilayer perceptron (MLP) for stacking; (7) Error assessment, with metrics including: mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and goodness of fit. .

[0007] Compared with the prior art, the present invention has the following beneficial effects:

[0008] 1. This invention proposes a method for identifying key events of high-speed aircraft based on infrared time series. By constructing a flight trajectory model and an infrared radiation model, it generates infrared time series data with physical constraints. On this basis, noise modeling and data augmentation methods are introduced to improve the diversity and authenticity of training data, thereby alleviating the problem of difficulty in obtaining real infrared observation data.

[0009] 2. In the event recognition stage, this invention combines a Temporal Convolutional Network (TCN) and an Extreme Gradient Boosting (XGBoost) model for key event detection. The TCN model effectively extracts long-term dependency features from infrared time series, while the XGBoost model captures local abrupt changes in the signal. Through a Stacking fusion strategy, the prediction results of multiple models are fused using a two-stage learning process, thereby fully utilizing the complementary advantages of different models to improve the accuracy and stability of key event recognition.

[0010] 3. Compared with traditional threshold detection methods, this invention can more accurately identify shutdown and deflagration events in the infrared radiation sequence of high-speed aircraft in complex noise environments, improving the robustness and reliability of infrared time series key event identification, and providing an effective method for infrared feature analysis and target status identification of high-speed aircraft.

[0011] A high-speed aircraft critical event identification device based on infrared radiation sequences is also provided, the device comprising: The flight trajectory model construction module is configured to establish the motion equation of the center of mass of a high-speed aircraft. It describes the motion state of the aircraft through forces such as thrust, aerodynamics, and gravity, and uses numerical integration to solve the motion equation to obtain the flight trajectory of the aircraft during the boost phase. The infrared radiation model building module is configured to construct an infrared radiation model of the tail flame based on flight trajectory data. It calculates the infrared radiation intensity through the tail flame geometry, temperature distribution, and observation distance, and generates an infrared radiation time series. The noise simulation module is configured to superimpose a Gaussian random number with a mean of zero onto each data point of the infrared radiation sequence, the standard deviation of which is adjusted by the signal amplitude, thereby simulating detection under different signal-to-noise ratio conditions; a periodic triangular wave, which is asymmetrical, is added to the signal, and perturbations are added to simulate periodic interference. The data augmentation module is configured to introduce augmentation methods for infrared radiation sequence data based on the base samples, constructing a time mapping function. A mapping from the original time t to the new time t' is established using cubic spline interpolation. The machine learning module is configured to identify events using an XGBoost regression model. To adapt to the input format of the XGBoost model, a sliding window is used to extract features and flatten the time series. To improve the model's generalization ability and prevent overfitting, key parameters of the XGBoost model are set. To fully learn the features of the infrared radiation sequence, a multi-layer dilated convolutional structure of TCN is used to increase the receptive field by gradually increasing the dilation coefficient, thereby capturing long-term dependencies in the infrared radiation sequence. During network training, to improve the model's generalization ability, a Dropout mechanism is introduced into the convolutional layers with a dropout rate of 0.2. The Adam optimization algorithm is used to optimize the model parameters with a learning rate of 0.001. The fusion module is configured to stack the results of different base models as input, and uses a multilayer perceptron (MLP) for stacking. The evaluation module is configured for error assessment, and the metrics include: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Goodness of Fit. . Attached Figure Description

[0012] Figure 1 The TCN structure is shown.

[0013] Figure 2 The Stacking structure is shown.

[0014] Figure 3 The MLP structure is shown.

[0015] Figure 4 An overall flowchart of the high-speed aircraft critical event identification method based on infrared radiation sequence according to the present invention is shown.

[0016] Figure 5 A flowchart of a specific embodiment of the high-speed aircraft critical event identification method based on infrared radiation sequence according to the present invention is shown.

[0017] Figure 6 The trajectory information is displayed.

[0018] Figure 7 The data shows augmented missile data with a shutdown point.

[0019] Figure 8 The data shows augmented missile data with shutdown and ignition points.

[0020] Figure 9 The XGBoost training process is shown.

[0021] Figure 10 The data shows augmented missile data with shutdown and ignition points.

[0022] Figure 11 This demonstrates the use of MLP for model fusion.

[0023] Figure 12 The output results of the MLP fusion training process are shown. Detailed Implementation

[0024] like Figure 5 As shown, this method for identifying critical events in high-speed aircraft based on infrared radiation sequences includes the following steps: (1) Establish the motion equation of the center of mass of the high-speed aircraft, describe the motion state of the aircraft through the forces such as thrust, aerodynamics and gravity, and solve the motion equation by numerical integration method to obtain the flight trajectory of the aircraft during the boost phase. (2) Based on flight trajectory data, construct an infrared radiation model of the tail flame, calculate the infrared radiation intensity through tail flame geometry, temperature distribution and observation distance, and generate an infrared radiation time series; (3) A Gaussian random number with a mean of zero is superimposed on each data point of the infrared radiation sequence, and its standard deviation is adjusted by the signal amplitude to simulate the detection under different signal-to-noise ratio conditions; a periodic triangular wave is added to the signal, which is an asymmetric triangular wave, and perturbation is added to simulate periodic interference. (4) Based on the basic sample, an augmentation method for infrared radiation sequence data is introduced to construct a time mapping function. A mapping from the original time t to the new time t' is established using cubic spline interpolation. (5) The XGBoost regression model is used to identify events. To adapt to the input format of the XGBoost model, a sliding window is used to extract features and flatten the time series. To improve the generalization ability of the model and prevent overfitting, the key parameters of the XGBoost model are set. To fully learn the features of the infrared radiation sequence, the multi-layer dilated convolution structure of TCN is used to gradually increase the receptive field by increasing the dilation coefficient and capture the long-term dependence in the infrared radiation sequence. During the network training process, in order to improve the generalization ability of the model, the Dropout mechanism is introduced in the convolutional layer with a dropout rate of 0.2. The Adam optimization algorithm is used to optimize the model parameters with a learning rate of 0.001. (6) Stacking takes the results of the base models of different models as input and uses a multilayer perceptron (MLP) for stacking; (7) Error assessment, with metrics including: mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and goodness of fit. .

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. This invention proposes a method for identifying key events of high-speed aircraft based on infrared time series. By constructing a flight trajectory model and an infrared radiation model, it generates infrared time series data with physical constraints. On this basis, noise modeling and data augmentation methods are introduced to improve the diversity and authenticity of training data, thereby alleviating the problem of difficulty in obtaining real infrared observation data.

[0027] 2. In the event recognition stage, this invention combines a Temporal Convolutional Network (TCN) and an Extreme Gradient Boosting (XGBoost) model for key event detection. The TCN model effectively extracts long-term dependency features from infrared time series, while the XGBoost model captures local abrupt changes in the signal. Through a Stacking fusion strategy, the prediction results of multiple models are fused using a two-stage learning process, thereby fully utilizing the complementary advantages of different models to improve the accuracy and stability of key event recognition.

[0028] 3. Compared with traditional threshold detection methods, this invention can more accurately identify shutdown and deflagration events in the infrared radiation sequence of high-speed aircraft in complex noise environments, improving the robustness and reliability of infrared time series key event identification, and providing an effective method for infrared feature analysis and target status identification of high-speed aircraft.

[0029] Preferably, step (1) is solved using formula (1): (1) Where: F is the thrust; For control; For aerodynamics; For gravity; This is the centrifugal inertial force caused by the Earth's rotation; This is the Coulomb inertial force; To obtain the trajectory, the centroid equation is solved using the fourth-order Runge-Kutta method: (2).

[0030] Preferably, in step (2), the infrared radiation time series is given by formula (3). (3) in: The equivalent projected area is calculated by the geometric projection algorithm; Real-time temperature of the exhaust plume of a high-speed aircraft; The distance from the high-speed aircraft to the observation point; It is the Stefan-Boltzmann constant; Emissivity of the tail flame; represents the detector band response coefficient.

[0031] Preferably, in step (2), the overall shape of the exhaust plume is approximated as a frustum, and a numerical method based on computer vision is used to accurately calculate the shape at different observation angles. The projected area of ​​the exhaust flame is calculated. To characterize the shutdown event, the shutdown point time is set to make the shutdown event controllable in advance. After the shutdown event occurs, the infrared intensity will show a characteristic of radiation decrease. The infrared radiation intensity is reduced by adjusting the exhaust flame temperature and geometric parameters. To characterize the deflagration event, the start time of the deflagration event is set to be randomly generated within a short interval after the first-stage shutdown. The deflagration duration is set to 10 time steps. To reflect the difference in deflagration intensity in different samples, a random pulse intensity parameter is introduced and superimposed on the original radiation sequence within the deflagration interval.

[0032] Preferably, in step (3), when adding two types of noise, Gaussian noise with varying intensity is first superimposed at different locations, then periodic triangular wave disturbances are added to the signal, and finally several Gaussian noises are superimposed.

[0033] Preferably, in step (4), the time mapping function is as follows: (4) Public It is a set of anchor points Construct a series of piecewise cubic polynomials. , making , And at each segment connection point, the function value, the first derivative, and the second derivative are guaranteed to be continuous; Infrared radiation is calculated using formula (5) to achieve amplitude scaling: (5).

[0034] Preferably, in step (5), the key parameters of the XGBoost model are set as follows: `max_depth` controls the maximum depth of a single decision tree, set to 6; `learning_rate` represents the contribution weight of each newly added tree to the final model, set to 0.05; regarding sample sampling, `subsample` and `colsample_bytree` control the random sampling ratio of training samples and features, respectively, both set to 0.8; the number of classes in the model, `num_class`, is set to 3, representing non-critical events, shutdown events, and explosion events; the random seed `random_state` is set to 42.

[0035] Preferably, in step (5), the TCN model is composed of an input layer, a feature extraction layer, and an output layer. First, the input layer receives single-channel infrared radiation time series data and inputs it into the TCN network for feature extraction. In the feature extraction stage, dilated convolution is used to gradually expand the receptive field of the convolution. The kernel size is set to 17, and the dilation coefficients are 1, 2, and 4 respectively. The corresponding number of channels is set to [32, 32, 32]. For sparse data, a loss function combining focal_loss and dice_loss is used to assist model training; the definition of focal_loss is Equation (6): (6) Where: L represents the loss; To predict the probability, if the sample belongs to class 1, then =p, if the sample belongs to class 0, then =1-p; This is the standard cross-entropy loss component; γ is the modulation factor, also known as the focusing parameter, which controls the degree of attention given to difficult-to-classify samples; As a weighting factor; dice_loss originates from image segmentation and is used to measure the degree of overlap between the predicted result and the true label. It is defined by formula (7): (7) in: Let be the predicted probability of the i-th sample; Let i be the true label of the i-th sample; As a smoothing factor, to prevent the denominator from being zero; The fusion method for these two loss functions is given by formula (8): (8) Where: L is the final loss function; The weight assigned to the diceloss; The weight assigned to focal loss; For discoreloss; For focal loss.

[0036] Preferably, in step (6), the MLP network contains three hidden layers, and the ReLU activation function is used in the hidden layers.

[0037] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When executed, the program includes the steps of the methods of the above embodiments. The storage medium can be ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a high-speed aircraft key event identification device based on infrared radiation sequences. This device includes: The flight trajectory model construction module is configured to establish the motion equation of the center of mass of a high-speed aircraft. It describes the motion state of the aircraft through forces such as thrust, aerodynamics, and gravity, and uses numerical integration to solve the motion equation to obtain the flight trajectory of the aircraft during the boost phase. The infrared radiation model building module is configured to construct an infrared radiation model of the tail flame based on flight trajectory data. It calculates the infrared radiation intensity through the tail flame geometry, temperature distribution, and observation distance, and generates an infrared radiation time series. The noise simulation module is configured to superimpose a Gaussian random number with a mean of zero onto each data point of the infrared radiation sequence, the standard deviation of which is adjusted by the signal amplitude, thereby simulating detection under different signal-to-noise ratio conditions; a periodic triangular wave, which is asymmetrical, is added to the signal, and perturbations are added to simulate periodic interference. The data augmentation module is configured to introduce augmentation methods for infrared radiation sequence data based on the base samples, constructing a time mapping function. A mapping from the original time t to the new time t' is established using cubic spline interpolation. The machine learning module is configured to identify events using an XGBoost regression model. To adapt to the input format of the XGBoost model, a sliding window is used to extract features and flatten the time series. To improve the model's generalization ability and prevent overfitting, key parameters of the XGBoost model are set. To fully learn the features of the infrared radiation sequence, a multi-layer dilated convolutional structure of TCN is used to increase the receptive field by gradually increasing the dilation coefficient, thereby capturing long-term dependencies in the infrared radiation sequence. During network training, to improve the model's generalization ability, a Dropout mechanism is introduced into the convolutional layers with a dropout rate of 0.2. The Adam optimization algorithm is used to optimize the model parameters with a learning rate of 0.001. The fusion module is configured to stack the results of different base models as input, and uses a multilayer perceptron (MLP) for stacking. The evaluation module is configured for error assessment, and the metrics include: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Goodness of Fit. .

[0038] The present invention will now be described in more detail.

[0039] This invention provides a method for identifying critical events in high-speed aircraft based on infrared radiation sequences, comprising the following steps:

[0040] 1. Establish a flight trajectory model

[0041] The motion equations of the center of mass of a high-speed aircraft are established. The motion state of the aircraft is described by the forces such as thrust, aerodynamics and gravity. The numerical integration method is used to solve the motion equations to obtain the flight trajectory of the aircraft during the boost phase.

[0042] , In the formula: F is the thrust; For control; For aerodynamics; For gravity; This is the centrifugal inertial force caused by the Earth's rotation; This is the Coulomb inertial force.

[0043] To obtain the trajectory, the centroid equation needs to be solved. This invention uses the fourth-order Runge-Kutta method. .

[0044] 2. Establish an infrared radiation model

[0045] Based on flight trajectory data, an infrared radiation model of the tail flame is constructed. The infrared radiation intensity is calculated by the tail flame geometry, temperature distribution and observation distance, and an infrared radiation time series is generated.

[0046] , In the formula: The equivalent projected area is calculated by the geometric projection algorithm; Real-time temperature of the exhaust plume of a high-speed aircraft; The distance from the high-speed aircraft to the observation point; It is the Stefan-Boltzmann constant; Emissivity of the tail flame; represents the detector band response coefficient.

[0047] The overall shape of the exhaust plume is approximated as a truncated cone, and a computer vision-based numerical method is used to accurately calculate the shape at different viewing angles. The projected area of ​​the exhaust flame.

[0048] To characterize the shutdown event, this invention makes the shutdown event pre-controllable by setting the engine shutdown point time. After the shutdown event occurs, the infrared intensity will exhibit a characteristic of decreasing radiation. The infrared radiation intensity is reduced by adjusting the exhaust flame temperature and geometric parameters.

[0049] To characterize deflagration events, this invention sets the start time of the deflagration event not to a fixed value, but to be randomly generated within a short interval after the first-level shutdown. The deflagration duration is set to 10 time steps. Simultaneously, to reflect the differences in deflagration intensity among different samples, a random pulse intensity parameter is introduced and superimposed on the original radiation sequence within the deflagration interval.

[0050] 3. Noise Modeling

[0051] To simulate noise conditions, this invention introduces Gaussian random noise and triangular wave perturbation into an ideal infrared radiation time series. A Gaussian random number with a mean of zero is superimposed on each data point in the infrared radiation sequence, and its standard deviation is adjusted by the signal amplitude, thereby simulating detection conditions under different signal-to-noise ratios. This processing can reflect the persistent and random fluctuations caused by factors such as sensor thermal noise and background fluctuations. In specific implementations, the noise is stronger in high-amplitude sections and weaker in low-amplitude sections, more closely resembling the noise characteristics of a real sensor.

[0052] The triangular wave uses an asymmetric triangular wave and adds perturbations to simulate periodic disturbances such as mechanical vibrations.

[0053] This invention adds two types of noise: first, Gaussian noise of varying intensity is superimposed at different locations; then, periodic triangular wave perturbations are added to the signal; and finally, Gaussian noise is superimposed several times. By adding Gaussian noise and triangular waves, the diversity and generalization ability of the data are improved.

[0054] 4. Data augmentation

[0055] To further construct the dataset for model training, this invention introduces an augmentation method for infrared radiation sequence data based on the basic samples.

[0056] To characterize the shutdown event, this invention makes the shutdown event pre-controllable by setting the engine shutdown point time. After the shutdown event occurs, the infrared intensity will exhibit a characteristic of decreasing radiation. The infrared radiation intensity is reduced by adjusting the exhaust flame temperature and geometric parameters.

[0057] To characterize deflagration events, this invention sets the start time of the deflagration event not to a fixed value, but to be randomly generated within a short interval after the first-level shutdown. The deflagration duration is set to 10 time steps. Simultaneously, to reflect the differences in deflagration intensity among different samples, a random pulse intensity parameter is introduced and superimposed on the original radiation sequence within the deflagration interval.

[0058] To achieve a nonlinear transformation of the time axis of an infrared radiation sequence, this invention constructs a time mapping function. A mapping from the original time t to the new time t' is established using cubic spline interpolation.

[0059] , Public It is a set of anchor points Construct a series of piecewise cubic polynomials. , making , And at each segment connection point, the function value is guaranteed to be continuous, the first derivative is continuous, and the second derivative is continuous.

[0060] Infrared radiation intensity is not only related to the target temperature, but also affected by factors such as observation distance, atmospheric absorption, and sensor gain, resulting in amplitude variations in the observed infrared intensity. To achieve amplitude scaling, the following formula is used: .

[0061] 5. Machine Learning

[0062] An XGBoost regression model was used to identify events.

[0063] The infrared time series has 1500 time steps, each containing normalized infrared radiation intensity. To adapt to the input format of the XGBoost model, a sliding window is used to extract features and flatten the time series. To improve the model's generalization ability and prevent overfitting, this invention sets key parameters for the XGBoost model.

[0064] To accommodate the input format of the XGBoost model, a sliding window is used to extract features and flatten the time series. To improve the model's generalization ability and prevent overfitting, this invention sets specific parameters for the XGBoost model.

[0065] The XGBoost parameter settings are shown in the table below.

[0066] Table 1 XGBoost Training Parameter Settings

[0067] `max_depth` controls the maximum depth of a single decision tree. A larger depth can improve the model's expressive power, but it may also increase the risk of overfitting. Therefore, this invention sets it to 6. `learning_rate` represents the contribution weight of each newly added tree in each round to the final model. A smaller learning rate can make the model training more stable. This invention sets its value to 0.05.

[0068] Regarding sample sampling, `subsample` and `colsample_bytree` are used to control the random sampling ratio of training samples and features, respectively, both set to 0.8 to enhance the model's generalization ability and reduce the risk of overfitting. Furthermore, the number of classes, `num_class`, is set to 3, representing non-critical events and two critical events: shutdown point and explosion point. To ensure the reproducibility of experimental results, a random seed, `random_state`, is set to 42 during training. These parameter configurations improve training stability.

[0069] To fully learn the characteristics of infrared radiation sequences, this invention utilizes the multi-layer dilated convolutional structure of TCN to gradually increase the receptive field by increasing the dilation coefficient, thereby capturing long-term dependencies in infrared radiation sequences.

[0070] The TCN network constructed in this invention consists of multiple residual blocks, each containing a one-dimensional convolutional layer, dropout, layer normalization, and a ReLU activation function. Each residual block alleviates the vanishing gradient problem during deep network training through residual connections. The structure of the TCN model in this invention is as follows: Figure 1 As shown.

[0071] The training parameters for the TCN model of this invention are shown in Table 2.

[0072] Table 2 TCN Training Parameter Settings

[0073] according to Figure 1 The TCN network structure shown in this invention uses infrared radiation time series as input, extracts multi-scale features from the time series through a multi-layer dilated convolutional network, and finally completes the identification of key events. The model consists of an input layer, a feature extraction layer, and an output layer.

[0074] First, the input layer receives single-channel infrared radiation time-series data and feeds it into the TCN network for feature extraction. During the feature extraction stage, this invention employs dilated convolution to progressively expand the receptive field. Specifically, the kernel size is set to 17, and the dilation coefficients are successively set to 1, 2, and 4, corresponding to a channel number of [32, 32, 32]. This layer-by-layer expansion of the receptive field structure can capture long-term dependencies in the infrared radiation sequence.

[0075] To improve the model's generalization ability during network training, a Dropout mechanism was introduced into the convolutional layers with a dropout rate of 0.2 to reduce overfitting. Model parameter optimization employed the Adam optimization algorithm with a learning rate of 0.001, ensuring stable convergence of the network during training.

[0076] Since key event points are relatively short compared to the total flight time, this invention creates a loss function that combines focal loss and dice loss to assist model training for sparse data.

[0077] First, focal loss is a function used in machine learning to calculate loss. Its key feature is its ability to address class imbalance by reducing the weights of easily classified samples and increasing the weights of difficult-to-classify samples, thus optimizing model training. The definition of focal loss is shown in the following equation: , In the formula: L represents the loss; To predict the probability, if the sample belongs to class 1, then =p, if the sample belongs to class 0, then =1-p; This is the standard cross-entropy loss component; γ is the modulation factor. When γ > 0, the loss of easily classified samples will be compressed, while the loss of difficult-to-classify samples will be larger. The larger the value of γ, the more obvious the effect of the modulation factor will be. γ is called the focusing parameter, which controls the degree of attention to difficult-to-classify samples. This is the weighting factor.

[0078] The focal loss function reduces the contribution of easily classified samples to the loss, allowing the model to focus more on the contribution of hard-to-classify samples during training, thus solving the class imbalance problem. Identifying key event points of high-speed aircraft is well-suited to the characteristics of focal loss. Key event points of high-speed aircraft constitute a very small minority of the aircraft's operational time, representing a sample imbalance problem. Traditional loss functions may be affected by a large number of non-key event points, resulting in insufficient attention to key events. Therefore, this invention uses the focal loss function.

[0079] Secondly, `dice_loss` originates from image segmentation and is used to measure the overlap between the predicted result and the true label. Its definition is shown in the equation: , In the formula: Let be the predicted probability of the i-th sample; Let i be the true label of the i-th sample; This is a smoothing factor to prevent the denominator from being zero.

[0080] The loss function `dice_loss` maximizes the overlap between the predicted results and the true labels. For key event point prediction, `dice_loss` effectively avoids biased or incomplete model predictions.

[0081] The method for fusing these two loss functions is shown in the following equation: , In the formula: L is the final loss function; The weight assigned to the diceloss; The weight assigned to focal loss; For discoreloss; This is a focal loss function. The weights can be flexibly adjusted based on the training results to achieve the best training outcome.

[0082] 6. Fusion Model

[0083] Stacking no longer relies on manually setting weights. Instead, it takes the results of different base models as input and trains secondary models to learn better mapping relationships. The structure of Stacking is as follows: Figure 2 As shown.

[0084] In the selection of the second-level model, this invention uses a multilayer perceptron (MLP) for stacking. The main reason is that MLP can theoretically approximate any continuous function, so it is suitable as a second-level fusion model for stacking to learn the nonlinear combination relationship between predictions from different models.

[0085] The MLP network constructed in this invention contains three hidden layers. The MLP structure constructed in this invention is as follows: Figure 3 As shown.

[0086] This invention uses the ReLU activation function in the hidden layer to enhance the model's nonlinear expressive power. This makes the MLP model more capable of learning the nonlinearity between the predictions of two models, thereby achieving more accurate identification of key events.

[0087] The parameters of the MLP model of this invention are shown in the table.

[0088] Table 3 MLP Training Parameter Settings

[0089] The evaluation indicators are as follows: Commonly used metrics include mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and goodness of fit (BFF). The calculation formulas are shown below. The smaller the error, the greater the goodness of fit, indicating that the model has higher accuracy and better prediction performance.

[0090] Mean Absolute Error (MAE): , Root Mean Square Error (RMSE): , Mean Absolute Percentage Error (MAPE): , Goodness of fit (R²): , in, For the true value, For predicted values, This represents the mean of the original sequence. Analyzing these metrics allows for a better understanding of the predictive model's performance, enabling adjustments and improvements to the model, thereby enhancing its recognition accuracy and performance.

[0091] Programming environment: Python, PyTorch.

[0092] Generate the trajectory, and visualize it as follows: Figure 6 As shown.

[0093] Infrared data generation, such as Figure 7 , 8 As shown.

[0094] Single model training, such as Figure 9 , 10 As shown.

[0095] Using MLP for model fusion, such as Figure 11 As shown.

[0096] The output results are as follows Figure 12 As shown. Figure 12 In the diagram: the horizontal axis represents time; the vertical axis represents infrared intensity; the black line represents the change in infrared radiation intensity over time; the red dot represents the time of the first-level shutdown point; the yellow dot represents the time of the deflagration point; the red dashed line represents the predicted time of the first-level shutdown point; and the yellow dashed line represents the predicted time of the deflagration point shutdown.

[0097] Results evaluation: Using RMSE, MAE, and R... 2 MAPE conducted an evaluation, as shown in Table 4.

[0098] Table 4

[0099] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for identifying critical events in high-speed aircraft based on infrared radiation sequences, characterized in that: The method includes the following steps: (1) Establish the motion equation of the center of mass of the high-speed aircraft, describe the motion state of the aircraft through the forces such as thrust, aerodynamics and gravity, and solve the motion equation by numerical integration method to obtain the flight trajectory of the aircraft during the boost phase. (2) Based on flight trajectory data, construct an infrared radiation model of the tail flame, calculate the infrared radiation intensity through tail flame geometry, temperature distribution and observation distance, and generate an infrared radiation time series; (3) A Gaussian random number with a mean of zero is superimposed on each data point of the infrared radiation sequence, and its standard deviation is adjusted by the signal amplitude to simulate the detection under different signal-to-noise ratio conditions; a periodic triangular wave is added to the signal, which is an asymmetric triangular wave, and perturbation is added to simulate periodic interference. (4) Based on the basic sample, an augmentation method for infrared radiation sequence data is introduced to construct a time mapping function. A mapping from the original time t to the new time t' is established using cubic spline interpolation. (5) The XGBoost regression model is used to identify events. To adapt to the input format of the XGBoost model, a sliding window is used to extract features and flatten the time series. To improve the generalization ability of the model and prevent overfitting, the key parameters of the XGBoost model are set. To fully learn the features of the infrared radiation sequence, the multi-layer dilated convolution structure of TCN is used to gradually increase the receptive field by increasing the dilation coefficient and capture the long-term dependence in the infrared radiation sequence. During the network training process, in order to improve the generalization ability of the model, the Dropout mechanism is introduced in the convolutional layer with a dropout rate of 0.

2. The Adam optimization algorithm is used to optimize the model parameters with a learning rate of 0.

001. (6) Stacking takes the results of the base models of different models as input and uses a multilayer perceptron (MLP) for stacking; (7) Error assessment, with metrics including: mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and goodness of fit. .

2. The method for identifying key events of high-speed aircraft based on infrared radiation sequences according to claim 1, characterized in that: The step (1) is solved using formula (1): （1） Where: F is the thrust; For control; For aerodynamics; For gravity; This is the centrifugal inertial force caused by the Earth's rotation; This is the Coulomb inertial force; To obtain the trajectory, the centroid equation is solved using the fourth-order Runge-Kutta method: （2）。 3. The method for identifying key events of high-speed aircraft based on infrared radiation sequences according to claim 2, characterized in that: In step (2), the infrared radiation time series is given by formula (3). （3） in: The equivalent projected area is calculated by the geometric projection algorithm; Real-time temperature of the exhaust plume of a high-speed aircraft; The distance from the high-speed aircraft to the observation point; It is the Stefan-Boltzmann constant; Emissivity of the tail flame; represents the detector band response coefficient.

4. The method for identifying key events of high-speed aircraft based on infrared radiation sequences according to claim 3, characterized in that: In step (2), the overall shape of the tail flame is approximated as a frustum, and a numerical method based on computer vision is used to accurately calculate the shape at different observation angles. The projected area of ​​the exhaust flame is calculated. To characterize the shutdown event, the shutdown point time is set to make the shutdown event controllable in advance. After the shutdown event occurs, the infrared intensity will show a characteristic of radiation decrease. The infrared radiation intensity is reduced by adjusting the exhaust flame temperature and geometric parameters. To characterize the deflagration event, the start time of the deflagration event is set to be randomly generated within a short interval after the first-stage shutdown. The deflagration duration is set to 10 time steps. To reflect the difference in deflagration intensity in different samples, a random pulse intensity parameter is introduced and superimposed on the original radiation sequence within the deflagration interval.

5. The method for identifying key events of high-speed aircraft based on infrared radiation sequences according to claim 4, characterized in that: In step (3), when adding two types of noise, first superimpose Gaussian noise of varying intensity at different locations, then add periodic triangular wave disturbances to the signal, and finally superimpose Gaussian noise several times.

6. The method for identifying key events of high-speed aircraft based on infrared radiation sequences according to claim 5, characterized in that: In step (4), the time mapping function is as follows: （4） Public It is a set of anchor points Construct a series of piecewise cubic polynomials. , making , And at each segment connection point, the function value, the first derivative, and the second derivative are guaranteed to be continuous; Infrared radiation is calculated using formula (5) to achieve amplitude scaling: （5）。 7. The method for identifying key events of high-speed aircraft based on infrared radiation sequences according to claim 6, characterized in that: In step (5), the key parameters of the XGBoost model are set as follows: `max_depth` controls the maximum depth of a single decision tree, set to 6; `learning_rate` represents the contribution weight of each newly added tree to the final model, set to 0.05; regarding sample sampling, `subsample` and `colsample_bytree` control the random sampling ratio of training samples and features, respectively, both set to 0.8; the number of classes in the model, `num_class`, is set to 3, representing non-critical events, shutdown events, and explosion events; the random seed `random_state` is set to 42.

8. The method for identifying key events of high-speed aircraft based on infrared radiation sequences according to claim 7, characterized in that: In step (5), the TCN model is composed of an input layer, a feature extraction layer, and an output layer. First, the input layer receives single-channel infrared radiation time series data and inputs it into the TCN network for feature extraction. In the feature extraction stage, dilated convolution is used to gradually expand the receptive field of the convolution. The kernel size is set to 17, and the dilation coefficients are 1, 2, and 4 respectively. The corresponding number of channels is set to [32, 32, 32]. For sparse data, a loss function combining focal_loss and dice_loss is used to assist model training; the definition of focal_loss is Equation (6): （6） Where: L represents the loss; To predict the probability, if the sample belongs to class 1, then =p, if the sample belongs to class 0, then =1-p; This is the standard cross-entropy loss component; γ is the modulation factor, also known as the focusing parameter, which controls the degree of attention given to difficult-to-classify samples; As a weighting factor; dice_loss originates from image segmentation and is used to measure the degree of overlap between the predicted result and the true label. It is defined by formula (7): （7） in: Let be the predicted probability of the i-th sample; Let i be the true label of the i-th sample; As a smoothing factor, to prevent the denominator from being zero; The fusion method for these two loss functions is given by formula (8): （8） Where: L is the final loss function; The weight assigned to the diceloss; The weight assigned to focal loss; For discoreloss; For focal loss.

9. The method for identifying key events of high-speed aircraft based on infrared radiation sequences according to claim 8, characterized in that: In step (6), the MLP network contains three hidden layers, and the ReLU activation function is used in the hidden layers.

10. A high-speed aircraft critical event identification device based on infrared radiation sequences, characterized in that: The device includes: The flight trajectory model construction module is configured to establish the motion equation of the center of mass of a high-speed aircraft. It describes the motion state of the aircraft through forces such as thrust, aerodynamics, and gravity, and uses numerical integration to solve the motion equation to obtain the flight trajectory of the aircraft during the boost phase. The infrared radiation model building module is configured to construct an infrared radiation model of the tail flame based on flight trajectory data. It calculates the infrared radiation intensity through the tail flame geometry, temperature distribution, and observation distance, and generates an infrared radiation time series. The noise simulation module is configured to superimpose a Gaussian random number with a mean of zero onto each data point of the infrared radiation sequence, the standard deviation of which is adjusted by the signal amplitude, thereby simulating detection under different signal-to-noise ratio conditions; a periodic triangular wave, which is asymmetrical, is added to the signal, and perturbations are added to simulate periodic interference. The data augmentation module is configured to introduce augmentation methods for infrared radiation sequence data based on the base samples, constructing a time mapping function. A mapping from the original time t to the new time t' is established using cubic spline interpolation. The machine learning module is configured to identify events using an XGBoost regression model. To adapt to the input format of the XGBoost model, a sliding window is used to extract features and flatten the time series. To improve the model's generalization ability and prevent overfitting, key parameters of the XGBoost model are set. To fully learn the features of the infrared radiation sequence, a multi-layer dilated convolutional structure of TCN is used to increase the receptive field by gradually increasing the dilation coefficient, thereby capturing long-term dependencies in the infrared radiation sequence. During network training, to improve the model's generalization ability, a Dropout mechanism is introduced into the convolutional layers with a dropout rate of 0.

2. The Adam optimization algorithm is used to optimize the model parameters with a learning rate of 0.

001. The fusion module is configured to stack the results of different base models as input, and uses a multilayer perceptron (MLP) for stacking. The evaluation module is configured for error assessment, and the metrics include: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Goodness of Fit. .

Images