Aircraft safety situation awareness improved model construction method based on adaptive knowledge distillation theory

Abstract

The invention discloses an aircraft safety situation awareness improved model construction method based on an adaptive knowledge distillation theory. The method comprises the following steps: firstly, training a teacher model by using terminal area ADS-B (Automatic Dependent Surveillance-Broadcast) data and a security situation level label correspondingly reflecting aircraft conflicts, and outputting a security situation level soft label; secondly, designing a student model, and learning the soft label and complex situation awareness ability of the teacher model; thirdly, the student model is trained under the guidance of the teacher model, and a self-adaptive dynamic temperature module and a terminal area operation safety rule constraint module are designed to guide training; and finally, designing a knowledge distillation total loss function, updating student model parameters by gradient descent, and obtaining a lightweight student model embedded with aviation safety domain knowledge. The method is suitable for constructing an improved model of a large aircraft safety situation awareness model in a terminal area, aviation safety domain knowledge is fully embedded, and an efficient, lightweight and safe practical model is obtained.

Description

Technical Field

[0001] This invention belongs to the field of intelligent air traffic management, specifically involving a method for constructing an improved aircraft safety situational awareness model based on adaptive knowledge distillation theory. Background Technology

[0002] Safety is the lifeline of the civil aviation industry, and efficiency is the relentless pursuit of civil aviation professionals. Knowledge distillation, an important model compression technique in machine learning, is applied to the optimization of aircraft safety situational awareness models. However, general distillation methods aim only to mimic the output of the teacher model, and their loss functions completely ignore the most core safety rules in the aviation field (such as the minimum interval standard). This can lead to distilled models making predictions that are close to the teacher's in performance but dangerous in decision-making. Furthermore, existing distillation techniques often use fixed temperature parameters to soften labels, which cannot adapt to the complex and variable uncertainties of aircraft operational data. Their loss functions only focus on reducing the output distribution differences between teacher and student models, completely ignoring the rationality of the output results in terms of aviation safety, potentially leading to catastrophic misjudgments by the compressed model.

[0003] Aircraft situational awareness models are often complex and require high computing power. In environments with limited computing power, deployment is difficult, ineffective, and the models have low practicality. Furthermore, this field demands extremely high model safety and reliability; existing distillation methods cannot simultaneously meet the requirements of efficiency, lightweight design, and safety, thus failing to effectively guarantee air traffic and the safety of people's lives and property. Therefore, this invention discloses a method for constructing an improved aircraft safety situational awareness model based on adaptive knowledge distillation theory. This method is applicable to the optimization and improvement of large aircraft safety situational awareness models in the terminal area, fully embedding knowledge from the aviation safety domain to obtain an efficient, lightweight, and safe practical model that meets the stringent requirements of real-time deployment and low power consumption. Summary of the Invention

[0004] Purpose of the Invention: Addressing the challenges of high computational requirements, difficult practical deployment, low model usability, and the inability of traditional distillation techniques to dynamically adapt to complex operational data in large-scale aircraft operation safety situation awareness models for large aircraft, this invention trains a high-performance teacher model and a more streamlined student model using training data containing ADS-B data of aircraft in the terminal area and corresponding safety situation level labels. It adjusts the KL divergence loss term by designing an adaptive dynamic temperature module combining Top-2 probability difference and safety weights, and designs terminal area operation safety rule constraints for calculating asymmetric penalty loss using a risk matrix. A total loss function for knowledge distillation is designed in conjunction with the standard loss of the real labels. Gradient descent is applied to the total loss function to update the student model parameters, resulting in an efficient, lightweight, and fully embedded safety knowledge-based practical model. This improves upon large and complex models, meeting the stringent requirements of real-time performance and low power consumption in practical deployment.

[0005] To achieve the above objectives, this invention provides a method for constructing an improved aircraft safety situational awareness model based on adaptive knowledge distillation theory, comprising the following steps:

[0006] Step S1: Train a teacher model using training data containing ADS-B data of aircraft in the terminal area and corresponding security situation level labels for the terminal area aircraft security situation awareness task, and output soft labels reflecting the security situation level of aircraft conflict.

[0007] Step S2: Design a student model whose input is a real-time or near-real-time sequence of aircraft operation features, learns the soft labels of the teacher model and inherits its ability to perceive complex situations.

[0008] Step S3: The student model is trained under the guidance of the teacher model. An adaptive dynamic temperature module that combines the Top-2 probability difference with the aircraft operation safety weights and a terminal area operation safety rule constraint module that calculates asymmetric penalty loss based on the risk matrix are designed to guide the distillation process and embed knowledge in the field of aviation safety.

[0009] Step 4: Integrate the outputs of the above modules, design the total loss function of knowledge distillation, minimize the total loss function through the gradient descent algorithm, update the student model parameters, and thus obtain a lightweight and practical model that fully embeds knowledge in the field of aviation safety.

[0010] Step S1 specifically includes:

[0011] A large, high-performance teacher model is trained using training data containing ADS-B data of aircraft in the terminal area and corresponding security situation level labels. This model is used for terminal area aircraft security situation awareness tasks, outputting soft labels reflecting the security situation level of aircraft conflicts. To accurately identify the security situation of aircraft in the terminal area and fully explore risk patterns, a deep clustering network is used as the teacher model. This model is trained end-to-end, simultaneously optimizing the high-quality extraction of aircraft operational features and the clustering of security levels. The two are collaboratively optimized through parameter sharing and gradient backpropagation, thereby achieving accurate and reliable discrimination of different situation levels under complex air situations. The teacher model constructs a deep encoder-decoder structure and a corresponding number of channels to form a high-dimensional, deep feature representation space and outputs a high-confidence soft clustering distribution, but its parameter count and computational complexity are also correspondingly high.

[0012] Based on the number of aircraft in the airspace within a certain time series, the safety situation level is divided into four levels: "normal", "concern", "dangerous" and "highly dangerous".

[0013] Using a security situation awareness matrix The following formula represents the airspace aircraft operational characteristics across all time periods:

[0014]

[0015] In the formula, Indicates time The perception matrix, Represents the characteristic number, For the number of aircraft in the airspace, This represents the total number of time series.

[0016] The security situation awareness matrix is ​​input into the teacher model, and the data is preprocessed first. Data normalization is performed on the security situation awareness matrix to eliminate differences in the dimensions of different features; one-hot encoding is then used to assign security situation levels.

[0017] The teacher model consists of two modules: a reconstruction and transformation layer and a clustering analysis layer. The teacher model loss function includes the teacher model reconstruction loss and the teacher model clustering loss. The loss function is minimized. The training model's ability to extract complex situational information and classify security situation levels using raw input data is as follows:

[0018]

[0019] In the formula, For the teacher model loss function, For teacher model reconstruction loss, For the clustering loss of the teacher model, These are the weight parameters corresponding to the teacher model loss.

[0020] Step S1.1: Rebuild the conversion layer

[0021] The reconstruction and transformation layer employs an autoencoder technique from an unsupervised learning model. By setting a specific number of encoders and decoders, along with the corresponding number of channels, it reconstructs and transforms the raw aircraft operation data, learning the latent knowledge within the data. The input data is transformed into an encoding matrix through a specific number of transformation layers. The decoder receives the encoding matrix and reconstructs the data to the greatest extent possible, ensuring consistency with the original input data. This allows the model to learn the spatiotemporal and risk characteristics of the aircraft's operational state, as shown in the following equation:

[0022]

[0023]

[0024] In the formula, For encoding features, To reconstruct features, For activation function, and For encoder weight parameters, and These are the decoder weight parameters.

[0025] For the continuous and discrete variables of the security situation awareness matrix, mean squared error (MSE) and cross-entropy (CE) are used as loss functions to minimize the reconstruction error. .

[0026]

[0027]

[0028]

[0029] In the formula, For continuous variable loss, For discrete variable loss, To rebuild the losses, For original continuous features, To reconstruct continuous features, For the original discrete features, To reconstruct discrete features, For the sample size, These are the weight parameters.

[0030] Step S1.2: Cluster Analysis Layer

[0031] The clustering analysis layer employs the Gaussian Mixture Model (GMM) algorithm from unsupervised learning to perform probabilistic soft partitioning of the aircraft operational features output by the reconstruction transformation layer. This layer acts as the bottleneck layer of the encoder, enabling a probabilistic description of the situation level and effectively characterizing the uncertainty of aircraft operational risks. The reconstructed features obtained from the reconstruction transformation layer... Evaluate the clustering effect of security status levels and minimize the clustering loss function. As shown in the following formula:

[0032]

[0033] In the formula, For the sample size, , To reconstruct features, , , These are the internal parameters of the GMM model.

[0034] Step S2 specifically includes:

[0035] Design a student model with fewer parameters and a more streamlined structure compared to the teacher model. Its input is a sequence of real-time or near-real-time aircraft operational features. The student model learns the teacher model's soft labels and inherits its ability to perceive complex situations. The student model is a lightweight, tailored version of the teacher model. It has the same feature extraction backbone network as the teacher model, but its network depth and width are compressed. The clustering accuracy is designed to approximate that of the teacher model, although there is usually a slight loss due to limitations in the student encoder's capabilities, but the speed is significantly improved. This isomorphic yet streamlined design facilitates knowledge transfer from intermediate feature maps and reduces the model's computational load.

[0036] The lightweight student model also includes two modules: a reconstruction and transformation layer and a clustering analysis layer. The student model loss function includes the student model reconstruction loss and the student model clustering loss. The goal is to minimize the student model loss function. As shown in the following formula:

[0037]

[0038] In the formula, For the student model loss function, For student model reconstruction loss, For student model clustering loss, These are the weight parameters corresponding to the student model loss.

[0039] Step S2.1: Rebuild the conversion layer

[0040] The reconstruction transformation layer of the student model uses an autoencoder network with both the number of layers and channels compressed, and its encoder layer number N stu With decoder layer number Mstu Less than the corresponding layer N of the teacher model tea With M tea .

[0041] For the continuous and discrete variables of the security situation awareness matrix, mean squared error (MSE) and cross-entropy (CE) are used as loss functions to minimize the reconstruction error. The calculation method is consistent with the teacher model.

[0042] Step S2.1: Cluster Analysis Layer

[0043] The clustering analysis layer of the student model inherits the clustering architecture of the teacher model, maintaining the same number of cluster centers. This ensures the consistency of the situational awareness hierarchy. The parameters of the clustering algorithm need to be randomly initialized and trained together with the encoder; the clustering loss... The clustering loss calculation method is consistent with that of the teacher model.

[0044] Step S3 specifically includes:

[0045] After constructing the teacher and student models, the student model was trained using an offline knowledge distillation strategy, while the teacher model parameters were frozen. To improve the efficiency of knowledge distillation training, a temperature parameter was introduced. As a modulating factor, the output probability distribution is softened, as shown in the following equation:

[0046]

[0047] In the formula, The probability output for each category, The logits output for each level, The temperature parameter is introduced. The higher the value, the greater the entropy of the soft label distribution, and the more the model will focus on negative labels.

[0048] (1) Adaptive dynamic temperature module

[0049] An adaptive dynamic temperature module is designed, combining Top-2 probability difference with safety weights. The Top-2 probability difference refers to the difference between the highest and second-highest probabilities in the model's predictions, used to quantify the model's certainty in making decisions between different categories. For each training sample, the dynamic temperature adjustment module calculates the temperature parameter based on the teacher model's prediction uncertainty for that sample. First, the baseline temperature value is calculated based on the Top-2 probability difference of the teacher model's prediction results. As shown in the following formula:

[0050]

[0051] In the formula, Based on the base temperature value, , These are the highest and second highest values ​​in the teacher model's situation level prediction probability. , Define the temperature fluctuation range for the adjustable hyperparameter. When >> That is, when the difference is close to 1, ≈ ;when ≈ That is, when the difference is close to 0, ≈ .

[0052] Then, the predicted safety level is amplified based on its corresponding safety weight, thereby generating a temperature parameter that adapts to sample uncertainty and safety weight. The following formula assigns different weight coefficients, from low to high, to the security status levels of "Normal," "Attention," "Danger," and "High Danger," respectively. .

[0053]

[0054] In the formula, Based on the base temperature value, For the final determined temperature parameters, The weighting coefficients for security status levels are determined by humans and can be calibrated through experience or evaluated by experts. This is a scaling factor that controls the overall impact of the safety weights.

[0055] For samples identified as "dangerous" or "highly dangerous," even if the teacher model has a high confidence level in its predictions, it will still assign higher temperature values ​​to these samples through an adaptive dynamic temperature module. By combining prediction uncertainty with preset safety weights, the teacher model outputs a smoother probability distribution, strengthening the student model's ability to learn subtle features of key risk scenarios, thereby improving its ability to identify high-risk situations.

[0056] Step S3.1:

[0057] During the training phase, temperature parameters were used. The output distributions of the teacher and student models are adjusted synchronously. The teacher model outputs soft labels for the security situation level, which are the probability distributions of each level output by the model's Softmax layer, supervising the training of the student model. The loss function is calculated based on the soft labels from the teacher model and the soft predictions output by the student model. Here, KL divergence is used to drive the student model to approximate the probability distribution characteristics of the teacher model, as shown in the following equation:

[0058]

[0059] In the formula, For the sample size, Let be the predicted distribution of the teacher model for the i-th sample. Let be the predicted distribution of the student model for the i-th sample.

[0060] Step S3.2:

[0061] During the inference phase, temperature parameters are set. The value is 1, meaning it is not used. The raw output logits of the student model are directly fed into the standard Softmax function, and the resulting probability distribution is the model's true confidence. The hard labels for the output situation level are then calculated, and the standard loss between the output and the true labels is calculated. Here, we use cross-entropy, as shown in the following formula:

[0062]

[0063] In the formula, For the true label of the sample, Hard labels predicted by the student model.

[0064] (2) Terminal area operation safety rule constraint module

[0065] The design proposes a terminal area operation safety rule constraint module based on risk matrix calculation of asymmetric penalty loss. The core step is to mathematize the prior knowledge and rules in the field of aviation safety, encode them into differentiable mathematical functions, and incorporate them as regularization terms into the total loss of knowledge distillation, thereby forcibly embedding safety constraints during the model optimization process.

[0066] By directly incorporating a risk assessment matrix from the security field into the loss function, asymmetric and differentiated penalties are applied to different types of misjudgments. First, a risk matrix is ​​defined. As shown in the following formula:

[0067]

[0068] In the formula, The true class for each sample is But it was predicted to be The risk costs depend on the research needs; This represents the number of situation level categories.

[0069] For the elements in the matrix Different risk costs are assigned based on different forecast scenarios. The values ​​are determined based on historical accident data, expert experience assessments, and air traffic control rules, and can be adjusted according to the specific airspace operational characteristics. Values ​​on the diagonal. The value is usually 0 because it's a correct prediction and therefore not penalized. Values ​​off-diagonal. This reflects asymmetric penalties, especially for predictions such as false positives and false negatives, which impose high risk costs to ensure that the model will be severely penalized when it makes a high-risk misjudgment.

[0070] Finally, the loss function of the safety rule constraint module based on the asymmetric penalty loss of the risk matrix is ​​calculated. By minimizing the loss function to fully consider the constraints of the terminal area's operational safety rules, as shown in the following equation:

[0071]

[0072] In the formula, For the sample size, For the first in this batch One sample, The total number of situation level categories, For risk matrix, The one-hot encoded vector of the real label represents the first... The true category of each sample; The probability distribution predicted by the student model.

[0073] Step S4 specifically includes:

[0074] By integrating the outputs of the above modules, a total loss function for knowledge distillation is designed. This total loss function is minimized using a gradient descent algorithm to update the student model parameters, resulting in a lightweight and practical model that fully incorporates knowledge from the aviation safety domain. This enables the optimization and improvement of a safety situational awareness model for large aircraft, meeting the stringent requirements of real-time performance and low power consumption for practical deployment.

[0075]

[0076] In the formula, Distillation of the total loss function for the student model. For safety constraint loss items, The KL divergence between the student model output and the teacher model soft label distribution is given. Output the standard loss between the student model and the true label. For temperature parameters, , , These are the weight parameters corresponding to different loss functions.

[0077] Compared to existing technologies, the advantages of this invention are as follows: This invention uses training data containing ADS-B data of aircraft in the terminal area and corresponding security situation level labels to train a high-performance teacher model and a more streamlined student model, respectively. By designing an adaptive dynamic temperature module that combines Top-2 probability difference with safety weights to adjust the KL divergence loss term, and designing a risk matrix to calculate asymmetric penalty loss for terminal area operation safety rules, and combining the standard loss with the real labels to design a knowledge distillation total loss function, the student model parameters are updated by gradient descent of the total loss function, resulting in an efficient, lightweight, and fully embedded safety knowledge practical model. This achieves the goal of improving large and complex models and meets the stringent requirements of real-time performance and low power consumption in actual deployment. Attached Figure Description

[0078] Figure 1 This is a flowchart illustrating the overall implementation of the present invention;

[0079] Figure 2 This is a structural diagram of the adaptive dynamic temperature module of the present invention;

[0080] Figure 3 This is a structural diagram of the terminal area operation safety rule constraint module of the present invention. Detailed Implementation

[0081] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0082] like Figure 1 As shown, this invention provides a method for constructing an improved aircraft safety situational awareness model based on adaptive knowledge distillation theory, comprising the following steps:

[0083] Step S1: Train a teacher model using terminal area ADS-B data and corresponding security situation level labels reflecting aircraft conflicts, and output soft labels for security situation levels.

[0084] To accurately identify the safety status of aircraft in the terminal area and fully explore risk patterns, this invention employs a deep clustering network as the teacher model. This model, through end-to-end training, simultaneously optimizes the high-quality extraction of aircraft operational features and the clustering of safety levels. These two aspects are collaboratively optimized through parameter sharing and gradient backpropagation, thereby achieving accurate and reliable discrimination of different status levels under complex air conditions. The teacher model constructs a deep encoder-decoder structure and a corresponding number of channels to form a high-dimensional, deep feature representation space and outputs a high-confidence soft clustering distribution; however, its parameter count and computational complexity are also correspondingly high.

[0085] Based on the number of aircraft in the airspace within a certain time series, the safety situation level is divided into four levels: "normal", "concern", "dangerous" and "highly dangerous".

[0086] Using a security situation awareness matrix The following formula represents the airspace aircraft operational characteristics across all time periods:

[0087]

[0088] In the formula, Indicates time The perception matrix, Represents the characteristic number, For the number of aircraft in the airspace, This represents the total number of time series.

[0089] The security situation awareness matrix is ​​input into the teacher model, and the data is preprocessed first. Data normalization is performed on the security situation awareness matrix to eliminate differences in the dimensions of different features; one-hot encoding is then used to assign security situation levels.

[0090] The teacher model consists of two modules: a reconstruction and transformation layer and a clustering analysis layer. The teacher model loss function includes the teacher model reconstruction loss and the teacher model clustering loss. The loss function is minimized. The training model's ability to extract complex situational information and classify security situation levels using raw input data is as follows:

[0091]

[0092] In the formula, For the teacher model loss function, For teacher model reconstruction loss, For the clustering loss of the teacher model, These are the weight parameters corresponding to the teacher model loss.

[0093] Step S1.1: Rebuild the conversion layer

[0094] The reconstruction and transformation layer employs an autoencoder technique from an unsupervised learning model. By setting a specific number of encoders and decoders, along with the corresponding number of channels, it reconstructs and transforms the raw aircraft operation data, learning the latent knowledge within the data. The input data is transformed into an encoding matrix through a specific number of transformation layers. The decoder receives the encoding matrix and reconstructs the data to the greatest extent possible, ensuring consistency with the original input data. This allows the model to learn the spatiotemporal and risk characteristics of the aircraft's operational state, as shown in the following equation:

[0095]

[0096]

[0097] In the formula, For encoding features, To reconstruct features, For activation function, and For encoder weight parameters, and These are the decoder weight parameters.

[0098] For the continuous and discrete variables of the security situation awareness matrix, mean squared error (MSE) and cross-entropy (CE) are used as loss functions to minimize the reconstruction error. .

[0099]

[0100]

[0101]

[0102] In the formula, For continuous variable loss, For discrete variable loss, To rebuild the losses, For original continuous features, To reconstruct continuous features, For the original discrete features, To reconstruct discrete features, For the sample size, These are the weight parameters.

[0103] Step S1.2: Cluster Analysis Layer

[0104] The clustering analysis layer employs the Gaussian Mixture Model (GMM) algorithm from unsupervised learning to perform probabilistic soft partitioning of the aircraft operational features output by the reconstruction transformation layer. This layer acts as the bottleneck layer of the encoder, enabling a probabilistic description of the situation level and effectively characterizing the uncertainty of aircraft operational risks. The reconstructed features obtained from the reconstruction transformation layer... Evaluate the clustering effect of security status levels and minimize the clustering loss function. As shown in the following formula:

[0105]

[0106] In the formula, For the sample size, The total number of clusters, i.e., the number of situation level categories. To reconstruct features, , , These are the internal parameters of the GMM model.

[0107] Step S2: Design a student model to learn the soft labels and complex situational awareness capabilities of the teacher model;

[0108] The student model is a lightweight version of the teacher model, featuring the same feature extraction backbone network, but with compressed network depth and width. While aiming to approximate the teacher model's clustering accuracy, it typically suffers slight performance loss due to limitations in the student encoder's capabilities, but its speed is significantly improved. This isomorphic yet streamlined design facilitates knowledge transfer from intermediate feature maps and reduces the model's computational load.

[0109] The student model also includes two modules: a reconstruction and transformation layer and a clustering analysis layer. The student model loss function comprises the student model reconstruction loss and the student model clustering loss. The goal is to minimize this student model loss function. As shown in the following formula:

[0110]

[0111] In the formula, For the student model loss function, For student model reconstruction loss, For student model clustering loss, These are the weight parameters corresponding to the student model loss.

[0112] Table 1: System Input Data

[0113]

[0114] Step S2.1: Rebuild the conversion layer

[0115] The reconstruction transformation layer of the student model uses an autoencoder network with both the number of layers and channels compressed, and its encoder layer number N stu With decoder layer number M stu Less than the corresponding layer N of the teacher model tea With M tea .

[0116] For the continuous and discrete variables of the security situation awareness matrix, mean squared error (MSE) and cross-entropy (CE) are used as loss functions to minimize the reconstruction error. The calculation method is consistent with the teacher model.

[0117] Step S2.2: Cluster Analysis Layer

[0118] The clustering analysis layer of the student model inherits the clustering architecture of the teacher model, maintaining the same number of cluster centers. This ensures the consistency of the situational awareness hierarchy. The parameters of the clustering algorithm need to be randomly initialized and trained together with the encoder; the clustering loss... The clustering loss calculation method is consistent with that of the teacher model.

[0119] Step S3: As Figure 2 and Figure 3 As shown, the student model is trained under the guidance of the teacher model, and the training is guided by an adaptive dynamic temperature module and a terminal area operation safety rule constraint module.

[0120] After constructing the teacher and student models, the student model was trained using an offline knowledge distillation strategy, while the teacher model parameters were frozen. To improve the efficiency of knowledge distillation training, a temperature parameter was introduced. As a modulating factor, the output probability distribution is softened, as shown in the following equation:

[0121]

[0122] In the formula, The probability output for each category, The logits output for each level, The temperature parameter is introduced. The higher the value, the greater the entropy of the soft label distribution, and the more the model will focus on negative labels.

[0123] The core innovations of this invention mainly include two aspects: firstly, an adaptive dynamic temperature module combining Top-2 probability difference and safety weight is designed; secondly, a terminal area operation safety rule constraint module based on risk matrix calculation of asymmetric penalty loss is designed.

[0124] (1) Adaptive dynamic temperature module

[0125] An adaptive dynamic temperature module is designed, combining Top-2 probability difference with safety weights. The Top-2 probability difference refers to the difference between the highest and second-highest probabilities in the model's predictions, used to quantify the model's certainty in making decisions between different categories. For each training sample, the dynamic temperature adjustment module calculates the temperature parameter based on the teacher model's prediction uncertainty for that sample. First, the baseline temperature value is calculated based on the Top-2 probability difference of the teacher model's prediction results. As shown in the following formula:

[0126]

[0127] In the formula, Based on the base temperature value, , These are the highest and second highest values ​​in the teacher model's situation level prediction probability. , Define the temperature fluctuation range for the adjustable hyperparameter. When >> That is, when the difference is close to 1, ≈ ;when ≈ That is, when the difference is close to 0, ≈ .

[0128] Then, the predicted safety level is amplified based on its corresponding safety weight, thereby generating a temperature parameter that adapts to sample uncertainty and safety weight. The following formula assigns different weight coefficients, from low to high, to the security status levels of "Normal," "Attention," "Danger," and "High Danger," respectively. .

[0129]

[0130] In the formula, Based on the base temperature value, For the final determined temperature parameters, The weighting coefficients for security status levels are determined by humans and can be calibrated through experience or evaluated by experts. This is a scaling factor that controls the overall impact of the safety weights.

[0131] For samples identified as "dangerous" or "highly dangerous," even if the teacher model has a high confidence level in its predictions, it still assigns higher temperature values ​​to these samples through an adaptive dynamic temperature module. By combining prediction uncertainty with preset safety weights, the teacher model outputs a smoother probability distribution, strengthening the student model's ability to learn subtle features of key risk scenarios, thereby improving its ability to identify high-risk situations.

[0132] Step S3.1: Training phase, using temperature parameters The output distributions of the teacher and student models are adjusted synchronously. The teacher model outputs soft labels for the security situation level, which are the probability distributions of each level output by the model's Softmax layer, supervising the training of the student model. The loss function is calculated based on the soft labels from the teacher model and the soft predictions output by the student model. Here, KL divergence is used to drive the student model to approximate the probability distribution characteristics of the teacher model, as shown in the following equation:

[0133]

[0134] In the formula, For the sample size, Let be the predicted distribution of the teacher model for the i-th sample. Let be the predicted distribution of the student model for the i-th sample.

[0135] Step S3.2: Inference stage, set temperature parameters The value is 1, meaning it is not used. The raw output logits of the student model are directly fed into the standard Softmax function, and the resulting probability distribution is the model's true confidence. The hard labels for the output situation level are then calculated, and the standard loss between the output and the true labels is calculated. Here, we use cross-entropy, as shown in the following formula:

[0136]

[0137] In the formula, For the true label of the sample, Hard labels predicted by the student model.

[0138] (2) Terminal area operation safety rule constraint module

[0139] The design proposes a terminal area operation safety rule constraint module based on risk matrix calculation of asymmetric penalty loss. The core step is to mathematize the prior knowledge and rules in the field of aviation safety, encode them into differentiable mathematical functions, and incorporate them as regularization terms into the total loss of knowledge distillation, thereby forcibly embedding safety constraints during the model optimization process.

[0140] By directly incorporating a risk assessment matrix from the security field into the loss function, asymmetric and differentiated penalties are applied to different types of misjudgments. First, a risk matrix is ​​defined. As shown in the following formula:

[0141]

[0142] In the formula, The true class for each sample is But it was predicted to be The risk costs depend on the research needs; This represents the number of situation level categories.

[0143] For the elements in the matrix Different risk costs are assigned based on different forecast scenarios. The values ​​are determined based on historical accident data, expert experience assessments, and air traffic control rules, and can be adjusted according to the specific airspace operational characteristics. Values ​​on the diagonal. The value is usually 0 because it's a correct prediction and therefore not penalized. Values ​​off-diagonal. This reflects asymmetric penalties, especially for predictions such as false positives and false negatives, which impose high risk costs to ensure that the model will be severely penalized when it makes a high-risk misjudgment.

[0144] Finally, the loss function of the safety rule constraint module with asymmetric penalty loss is calculated based on the risk matrix. By minimizing the loss function to fully consider the constraints of the terminal area's operational safety rules, as shown in the following equation:

[0145]

[0146] In the formula, For the sample size, For the first in this batch One sample, For the number of situation level categories, For risk matrix, The one-hot encoded vector of the real label represents the first... The true category of each sample; The probability distribution predicted by the student model.

[0147] Step 4: Design the total loss function for knowledge distillation, update the student model parameters using gradient descent, and obtain a lightweight student model with embedded aviation safety domain knowledge.

[0148] By integrating the outputs of the above modules, a total loss function for knowledge distillation is designed. This total loss function is minimized using a gradient descent algorithm to update the student model parameters, resulting in a lightweight and practical model that fully incorporates knowledge from the aviation safety domain. This enables the optimization and improvement of a safety situational awareness model for large aircraft, meeting the stringent requirements of real-time performance and low power consumption for practical deployment.

[0149]

[0150] In the formula, Distillation of the total loss function for the student model. For safety constraint loss items, The KL divergence between the student model output and the teacher model soft label distribution is given. Output the standard loss between the student model and the true label. For temperature parameters, , , These are the weight parameters corresponding to different loss functions.

[0151] Table 2: Output Evaluation Results Table

[0152]

[0153] This invention uses training data containing ADS-B data of aircraft in the terminal area and corresponding safety status level labels to train a high-performance teacher model and a more streamlined student model. It adjusts the KL divergence loss term by designing an adaptive dynamic temperature module that combines Top-2 probability difference with safety weights, and designs terminal area operation safety rule constraints to calculate asymmetric penalty loss using a risk matrix. It also designs a knowledge distillation total loss function based on the standard loss of the real labels, and updates the student model parameters using gradient descent on the total loss function. This results in a practical model that is efficient, lightweight, and fully embeds safety knowledge, achieving the goal of improving large and complex models and meeting the stringent requirements of real-time performance and low power consumption in actual deployment.

Claims

1. A method for constructing an improved aircraft safety situational awareness model based on adaptive knowledge distillation theory, characterized in that, Includes the following steps: Step S1: Train a teacher model using training data containing ADS-B data of aircraft in the terminal area and corresponding security situation level labels for the terminal area aircraft security situation awareness task, and output soft labels reflecting the security situation level of aircraft conflict. Step S2: Design a student model whose input is a real-time or near-real-time sequence of aircraft operation features, learns the soft labels of the teacher model and inherits its ability to perceive complex situations. Step S3: The student model is trained under the guidance of the teacher model. An adaptive dynamic temperature module that combines the Top-2 probability difference with the aircraft operation safety weights and a terminal area operation safety rule constraint module that calculates asymmetric penalty loss based on the risk matrix are designed to guide the distillation process and embed knowledge from the aviation safety domain. Step S4: Integrate the outputs of the above modules, design the total loss function of knowledge distillation, minimize the total loss function through the gradient descent algorithm, update the student model parameters, and thus obtain a lightweight and practical model with embedded aviation safety domain knowledge.

2. The method for constructing an improved aircraft safety situational awareness model based on adaptive knowledge distillation theory according to claim 1, characterized in that, In step S1: To accurately identify the safety status of aircraft in the terminal area and fully explore risk patterns, a deep clustering network is used as the teacher model. This model is trained end-to-end, simultaneously optimizing the high-quality extraction of aircraft operational features and the clustering of safety levels. The two are collaboratively optimized through parameter sharing and gradient backpropagation, thereby achieving accurate and reliable discrimination of different status levels under complex air conditions. The teacher model forms a high-dimensional, deep feature representation space by constructing a deep encoder-decoder structure and corresponding number of channels, and outputs a high-confidence soft clustering distribution. Based on the number of aircraft in the airspace within a certain time series, the security situation level is divided into four levels: "normal", "concern", "dangerous" and "highly dangerous". Using a security situation awareness matrix The following formula represents the airspace aircraft operational characteristics across all time periods: ； In the formula, Indicates time The perception matrix, Represents the characteristic number, For the number of aircraft in the airspace, Indicates the total number of timing sequences; The security situation awareness matrix is ​​input into the teacher model, and the data is preprocessed first; the security situation awareness matrix is ​​normalized to eliminate differences in the dimensions of different features; and the security situation level is one-hot encoded. The teacher model consists of two modules: a reconstruction and transformation layer and a clustering analysis layer. The teacher model loss function includes the teacher model reconstruction loss and the teacher model clustering loss. The loss function is minimized. The training model's ability to extract complex situational information and classify security situation levels using raw input data is as follows: ； In the formula, For the teacher model loss function, For teacher model reconstruction loss, For the clustering loss of the teacher model, These are the weight parameters corresponding to the teacher model loss. (1) Reconstructing the conversion layer The reconstruction and transformation layer employs an autoencoder technique from an unsupervised learning model. By setting a specific number of encoders and decoders and corresponding channels, it reconstructs and transforms the original aircraft operation data, learning the latent knowledge within the data. The input data is transformed into an encoding matrix through a specific number of transformation layers. The decoder receives the encoding matrix and reconstructs the data to the greatest extent possible, ensuring consistency with the original input data. This enables the model to learn the spatiotemporal and risk characteristics of the aircraft's operational state, as shown in the following equation: ； ； In the formula, For encoding features, To reconstruct features, For activation function, and For encoder weight parameters, and These are the decoder weight parameters; For the continuous and discrete variables of the security situation awareness matrix, mean squared error (MSE) and cross-entropy (CE) are used as loss functions to minimize the reconstruction error. ; ； ； ； In the formula, For continuous variable loss, For discrete variable loss, To rebuild the losses, For original continuous features, To reconstruct continuous features, For the original discrete features, To reconstruct discrete features, For the sample size, These are weight parameters; (2) Cluster analysis layer The clustering analysis layer employs the Gaussian Mixture Model (GMM) algorithm from unsupervised learning to perform probabilistic soft partitioning of the aircraft operational features output by the reconstruction and transformation layer. This layer acts as the bottleneck layer of the encoder, enabling a probabilistic description of the situation level and effectively characterizing the uncertainty of aircraft operational risks. Based on the reconstructed features obtained from the reconstruction and transformation layer... Evaluate the clustering effect of security status levels and minimize the clustering loss function. As shown in the following formula: ； In the formula, For the sample size, The total number of clusters, i.e., the number of situation level categories. To reconstruct features, , , These are the internal parameters of the GMM model.

3. The method for constructing an improved aircraft safety situational awareness model based on adaptive knowledge distillation theory according to claim 1, characterized in that, In step S2: Design a student model, a lightweight version tailored to the teacher model. It possesses the same feature extraction backbone network as the teacher model, but with compressed network depth and width. The clustering accuracy is designed to approximate that of the teacher model. The input is a real-time or near-real-time sequence of aircraft operational features. It also includes two modules: a reconstruction and transformation layer and a clustering analysis layer. The student model's loss function comprises the student model reconstruction loss and the student model clustering loss. The goal is to minimize this student model loss function. As shown in the following formula: ； In the formula, For the student model loss function, For student model reconstruction loss, For student model clustering loss, The weight parameters corresponding to the student model loss; (1) Reconstructing the conversion layer The reconstruction transformation layer of the student model uses an autoencoder network with both the number of layers and channels compressed, and its encoder layer number N stu With decoder layer number M stu Less than the corresponding layer N of the teacher model tea With M tea ; For the continuous and discrete variables of the security situation awareness matrix, mean squared error (MSE) and cross-entropy (CE) are used as loss functions to minimize the reconstruction error. The calculation method is consistent with the teacher model; (2) Cluster analysis layer The clustering analysis layer of the student model inherits the clustering architecture of the teacher model, maintaining the same number of cluster centers. To ensure the consistency of the situational awareness hierarchy; the parameters of the clustering algorithm need to be randomly initialized and trained together with the encoder; clustering loss The clustering loss calculation method is consistent with that of the teacher model.

4. The method for constructing an improved aircraft safety situational awareness model based on adaptive knowledge distillation theory according to claim 1, characterized in that, In step S3: After constructing the teacher and student models, an offline knowledge distillation strategy was used to train the student model, while the teacher model parameters were frozen. To improve the efficiency of knowledge distillation training, a temperature parameter was introduced. As a modulating factor, the output probability distribution is softened, as shown in the following equation: ； In the formula, The probability output for each category, The logits output for each level, For the temperature parameters introduced; The higher the value, the greater the entropy of the soft label distribution, and the more the model will focus on negative labels; (1) Adaptive dynamic temperature module An adaptive dynamic temperature module is designed, combining Top-2 probability difference with safety weights. The Top-2 probability difference refers to the difference between the highest and second-highest probabilities in the model's predictions, used to quantify the certainty of the model's decision-making among different categories. For each training sample, the dynamic temperature adjustment module calculates the temperature parameter based on the teacher model's prediction uncertainty for that sample. First, determine the Top-2 probability difference based on the teacher model prediction results, and then calculate the baseline temperature value. As shown in the following formula: ； In the formula, Based on the base temperature value, , These are the highest and second highest probabilities in the teacher's situation level prediction model. , Define the temperature fluctuation range for adjustable hyperparameters; when >> That is, when the difference is close to 1, ≈ ;when ≈ That is, when the difference is close to 0, ≈ ; Then, the predicted safety level is amplified based on its corresponding safety weight, thereby generating a temperature parameter that adapts to both sample uncertainty and safety weight. As shown in the following formula, different weight coefficients from low to high are assigned to the security status levels "Normal", "Attention", "Danger", and "High Danger", respectively. ; ； In the formula, Based on the base temperature value, For the final determined temperature parameters, The weighting coefficients for security status levels are determined by humans and can be calibrated through experience or evaluated by experts. This is a scaling factor that controls the overall impact of the safety weights. For samples identified as "dangerous" or "highly dangerous", even if the teacher model has a high confidence in its prediction results, it still assigns a higher temperature value to such samples through an adaptive dynamic temperature module; by combining prediction uncertainty with preset safety weights, the teacher model outputs a smoother probability distribution, strengthens the student model's learning of the subtle features of key risk scenarios, and thus improves its ability to identify high-risk situations. Step S3.1: Training phase, using temperature parameters The output distributions of the teacher and student models are adjusted synchronously. The teacher model outputs soft labels for the security situation level, which are the probability distributions of each level output by the model's Softmax layer, thus supervising the training of the student model. The loss function is calculated based on the soft labels from the teacher model and the soft predictions output by the student model. Here, KL divergence is used to drive the student model to approximate the probability distribution characteristics of the teacher model, as shown in the following equation: ； In the formula, For the sample size, Let be the predicted distribution of the teacher model for the i-th sample. Let be the predicted distribution of the student model for the i-th sample; Step S3.2: Inference stage, set temperature parameters It is a constant of 1, meaning it is not used. The raw output logits of the student model are directly fed into the standard Softmax function, and the resulting probability distribution is the model's true confidence. The hard labels for the output situation level are then calculated, and the standard loss between the output and the true labels is calculated. Here, we use cross-entropy, as shown in the following formula: ； In the formula, For the true label of the sample, Hard labels predicted by the student model; (2) Terminal area operation safety rule constraint module The design proposes a terminal area operation safety rule constraint module based on risk matrix calculation of asymmetric penalty loss. The core step is to mathematize the prior knowledge and rules in the field of aviation safety, encode them into differentiable mathematical functions, and incorporate them as regularization terms into the total loss of knowledge distillation, thereby forcibly embedding safety constraints during the model optimization process. By directly incorporating a risk assessment matrix from the security field into the loss function, asymmetric and differentiated penalties are applied to different types of misjudgments. First, a risk matrix is ​​defined. As shown in the following formula: ； In the formula, The true class for each sample is But it was predicted to be The risk costs depend on the research needs; The number of situation level categories; For the elements in the matrix Different risk costs are assigned based on different forecast scenarios. The values ​​are determined based on historical accident data, expert experience assessments, and air traffic control rules, and can be adjusted according to the specific airspace operational characteristics; the values ​​on the diagonal... It is usually 0 because the prediction is correct and there is no penalty; values ​​off-diagonal. This reflects asymmetric penalties, especially for predictions such as false positives and false negatives, which impose high risk costs to ensure that the model will be severely penalized when it makes a high-risk misjudgment. Finally, the loss function of the safety rule constraint module with asymmetric penalty loss is calculated based on the risk matrix. By minimizing the loss function to fully consider the constraints of the terminal area's operational safety rules, as shown in the following equation: ； In the formula, For the sample size, For the first in this batch One sample, For the number of situation level categories, For risk matrix, The one-hot encoded vector of the real label represents the first... The true category of each sample; The probability distribution predicted by the student model.

5. The method for constructing an improved aircraft safety situational awareness model based on adaptive knowledge distillation theory according to claim 1, characterized in that, In step S4: Design a total loss function for knowledge distillation to optimize and improve the safety situation awareness model for large aircraft, meeting the stringent requirements of real-time performance and low power consumption for actual deployment; ； In the formula, Distillation of the total loss function for the student model. For safety constraint loss items, The KL divergence between the student model output and the teacher model soft label distribution is given. Output the standard loss between the student model and the true label. For temperature parameters, , , These are the weight parameters corresponding to different loss functions.

Images