Pilot type continuous identification method based on topological information incremental neural framework
By employing a dynamic topology adaptation mechanism and combined loss function optimization, the small sample size problem and continuous update requirement in pilot type recognition are addressed, achieving efficient and accurate pilot type recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-06
- Publication Date
- 2026-04-03
AI Technical Summary
Pilot type recognition faces the challenges of small sample sizes and the need for continuous updates. Existing incremental learning methods are insufficient in terms of computational resources and adaptability, making it difficult to meet real-time recognition requirements.
We employ an incremental learning method based on topology networks. Through the dynamic topology adaptation mechanism of Neural Gas Network (NG), combined with the optimization of Elastic Weight Consolidation (EWC) loss, dynamic topology preservation loss, prototype contrast loss and classification loss, we achieve adaptation to new categories and retention of old knowledge.
In scenarios where data is scarce and needs to be continuously updated, efficient identification of pilot types is achieved, avoiding overfitting due to small samples and forgetting of old knowledge, thus improving the accuracy and adaptability of identification.
Smart Images
Figure CN121786442A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of aviation data analysis, specifically involving an incremental learning method for pilot type identification. Background Technology
[0002] Pilot type identification is crucial for personalized decision-making and achieving cognitive consistency between humans and machines. Pilot type identification can be achieved through flight data, directly acquiring flight status (latitude, longitude, altitude, vacuum speed, etc.) and operational parameters (pitch angle, roll angle, etc.) recorded by existing avionics systems (such as FDR / QAR). This requires no additional sensors and does not affect the pilot's normal operations. The status and operational features extracted from flight data can characterize the behavioral tendencies of a class of pilots, eliminating the need to model every pilot and resulting in lower modeling costs. During flight, individual differences among pilots are generally only apparent at key decision points, and only data before and after these key decision points can serve as valid data for the pilot type identification model. Due to the limited number of key decision points and the small amount of valid data in a single flight, pilot type identification is a typical small-sample problem. Furthermore, when new pilots join and do not belong to existing pilot categories, incremental learning of the pilot type identification model is necessary to ensure the system's continuous personalized matching capabilities.
[0003] Currently, incremental learning methods include memory replay, knowledge distillation, meta-learning, and topological networks. Rebuffi et al. proposed combining the Nearest Class Mean (NCM) classifier with sample replay to reduce computational complexity, and using cross-entropy loss and distillation loss to balance new and old knowledge, thus constructing an efficient and scalable incremental learning framework, which systematically solved the class increment problem for the first time [1]. Castro et al. proposed an end-to-end training method that jointly optimizes the feature extractor and classifier, using cross-entropy loss to learn new classes, using distillation loss to retain old class knowledge, balancing the gradient updates of new and old samples to alleviate the forgetting problem, and solving the class imbalance problem by sampling a subset of the new class with the same size as the old class example set, thereby improving model stability [2]. Prabhu et al. proposed a general incremental learning framework without assumption constraints, challenging the effectiveness of existing incremental learning methods by greedy sample storage and retraining during testing, revealing the possible biases of current evaluation metrics, and advocating for more rigorous experimental design methods [3]. Li et al. proposed training the model only with new category data and using knowledge distillation to retain old category knowledge, so that new and old samples share part of the feature space and avoid the need to store historical data [4]. Zhao et al. proposed a few-sample incremental learning framework based on class-aware bilateral distillation (CABD) [5]. Through the collaborative optimization of teacher network and student network, combined with feature space alignment and class prototype management, the problems of overfitting and catastrophic forgetting are alleviated. Chi et al. proposed an incremental learning framework based on the "meta-training-meta-testing" paradigm [6]. It aims to learn generalizable incremental strategies from the base class, without storing old class samples, and only relying on feature space optimization, which significantly reduces storage overhead. Zhang et al. proposed a continuously evolving classifier (CEC) [7]. The classifier parameters are dynamically adjusted through graph neural network (GNN), which solves the problem of classifier rigidity in traditional incremental learning. That is, the fixed classifier is difficult to adapt to the dynamic relationship between new and old categories, resulting in a sharp drop in performance in the few-sample scenario. Tao et al. proposed an incremental learning method based on topological networks [8], which transforms the relationship between new categories into topological relationships, uses Ancher Loss to solve the catastrophic forgetting problem, makes the new model as close as possible to the old model, uses Min-Max Loss to avoid overfitting caused by small samples, makes the model more discriminative, narrows the distance between samples of the same class, and widens the distance between samples of different classes.
[0004] In summary, memory replay methods suffer from high storage overhead and insufficient sample representativeness with small sample sizes; knowledge distillation methods exhibit significant class imbalance, struggle to balance performance between new and old classes with small sample sizes, and carry a high risk of overfitting; meta-learning methods rely on numerous meta-training tasks, demanding significant computational resources, making them unsuitable for pilot type recognition scenarios. Topological networks, however, possess self-organizing capabilities, dynamically adjusting neuron distribution through unsupervised competitive learning to automatically adapt to new classes without requiring predefined network structures; they can construct stable class prototypes with only a small number of samples, effectively avoiding overfitting; and the number of parameters is linearly related to the number of classes, resulting in lower computational costs, making them suitable for real-time incremental systems in pilot type recognition. Therefore, topological networks, such as neural gas networks, significantly outperform other methods in terms of dynamic adaptability, small sample efficiency, and real-time performance, making them particularly suitable for pilot type recognition scenarios where data is scarce and continuous updates are required. Summary of the Invention
[0005] Objective: To address the problem of pilot type identification in scenarios with scarce data and the need for continuous updates, this invention proposes a continuous pilot type identification method based on an incremental neural framework of topological information. In the feature extraction stage, spatial situational features of different pilots are extracted based on prior knowledge. In the topological network layer, a dynamic topology adaptation mechanism is designed; when the distance between a new category sample and an existing prototype exceeds a threshold, a new neuron is added. A joint training loss function is constructed, fusing Elastic Weight Consolidation (EWC) loss, dynamic topology preservation loss, prototype contrast loss, and classification loss. This method effectively solves the problems of overfitting with small samples and forgetting of old knowledge, and shows promising application prospects in pilot type identification scenarios with scarce data and the need for continuous updates. Technical solution: A persistent pilot type recognition method based on an incremental neural framework of topological information includes the following steps: S1, obtain t Group pilot category labeled dataset , ; S2, from the current data Extracting relative spatial situation features and local features, and calculating the statistical features of all the above features within a time window, the resulting feature space is denoted as […]. ; S3, when i When =1, enter S4, when i If the value is greater than 1, proceed to step S5; S4, construct a... Neural gas network of nodes , It describes the network topology. ( ) represents a node in the network. Let the set of edges represent the adjacency relationships between nodes; calculate each feature. Distance to each node Each feature Assign to the nearest node, and allocate the feature space. Divide the regions into unconnected areas and proceed to step S7; S5, Design an adaptive threshold to detect the current feature. If the distribution of nodes is significantly different from the existing node distribution, and the difference is greater than the threshold, then a new category has emerged, and new nodes need to be added to dynamically expand the NG network, proceeding to step S6; if the difference is less than the threshold, proceed to step S7. S6, Design the combination loss function Using the Adam algorithm to Optimize to obtain the current topology of the NG network and the optimal parameters of the classifier, then proceed to step S8; S7, Constructing the classification loss Using the Adam algorithm to Optimize the classifier to obtain its optimal parameters; S8, using current data Extracted features Calculate features Distance to each NG node Input the data into the classifier to obtain the pilot's category.
[0006] Furthermore, the pilot category annotation dataset The data was exported from ACMI format to Excel format using Tacview software at a sampling frequency of 20 Hz. The exported parameters included flight status parameters such as longitude, latitude, and altitude of both sides, as well as operational status parameters such as roll and pitch angles, totaling 61 dimensions. Taking the pilot's evasion of enemy missiles as the decision point, data was extracted 2.5 seconds before and after that moment, and the pilot type of this set of data was labeled as pilot category label data. Therefore, the data time window was 5 seconds.
[0007] Furthermore, based on current data Extract relative spatial situation features, including relative altitude, relative distance, azimuth, elevation angle, and proximity rate; (1) The relative height is calculated using the following formula:
[0008] This refers to the altitude of the machine. For the first The altitude of the enemy aircraft For this machine and the first The altitude difference between the enemy aircraft; (2) The relative distance is calculated using the following formula:
[0009] These are the local machine's geocentric rectangular coordinates. For the first The geocentric rectangular coordinates of the enemy aircraft For this machine and the first The spatial distance between enemy aircraft; (3) The azimuth angle is calculated using the following formula:
[0010] Where E and N are obtained by transforming relative vectors to the local coordinate system of the machine:
[0011] in , indicating the first The enemy aircraft's coordinates relative to this aircraft; East E extends eastward along the latitude circle, North N extends northward along the longitude line, and U is perpendicular to the ground upwards. Longitude Latitude; (4) The pitch angle is calculated using the following formula:
[0012] (5) The proximity rate is calculated using the following formula:
[0013] in, , , Let E be the eastward component of the velocity of the k-th enemy aircraft. The E component of my machine's speed; Let N be the north component of the velocity of the k-th enemy aircraft. The N component represents the speed of my machine.
[0014] Furthermore, based on current data Extracting native features, including energy characteristics, turning radius, and angular velocity: (1) The formula for calculating energy characteristics is as follows:
[0015] in For the machine's speed, This refers to the altitude of the machine. It is the acceleration due to gravity; (2) The turning radius is denoted as R, which can be directly obtained from the dataset. Read from; (3) The formula for calculating angular velocity is as follows: .
[0016] Furthermore, the statistical characteristics of all the above features within the time window are calculated, including: A total of 100 sampling points were collected, with a time window of 5 seconds and a sampling frequency of 20 Hz. Calculate the statistical characteristics of all the above features within the time window: The formula for calculating the mean is: ; The formula for calculating standard deviation is: ; The formula for calculating skewness is: ; The formula for calculating kurtosis is: ; in This represents a specific feature at a given sampling point. This represents the mean of a certain feature within a time window. This represents the standard deviation of a certain feature within a time window. This indicates the skewness of a feature within a time window. This represents the kurtosis of a feature within a time window, where n is the number of samples within a time window.
[0017] Furthermore, when i When =1, construct a... Neural gas network of nodes , It describes the network topology. ( ) represents a node in the network. Let the set of edges represent the adjacency relationships between nodes; calculate each feature. Distance to each node ,
[0018] in For each node's centroid, sort the distances from it in ascending order to obtain the following arrangement of nodes:
[0019] So, for each node Its center of gravity Updated to :
[0020] in For learning rate, The decay function, The updated centroid is a constant. Closer to features ; Features nearest node and the second nearest node Update the node's "age" and edge connections; The "age" and the updates to the connected edges are as follows: ,
[0021] node The "age" and the updates to the connected edges are as follows: ,
[0022] and The connection between them is made by letting and corresponding age To create or maintain; if the neuron is or The "age" is reset to 0; for other neurons, the age is increased by 1, and neurons with higher ages are deleted; for other edges, if... Exceeding the maximum age , The edge was deleted. Through the above operations, each feature Assign to the nearest node, and allocate the feature space. Divide into unconnected regions.
[0023] Furthermore, when i When the threshold is greater than 1, design an adaptive threshold:
[0024] in , Historical characteristics The average error and standard deviation of the distribution of historical nodes; Input current feature Find the nearest node Its center of gravity is Detect whether the current feature is significantly different from the distribution of existing nodes. If:
[0025] Then add a new neuron.
[0026] Furthermore, wheni When the value is greater than 1, design the combination loss function. Using the Adam algorithm
[15] to... Optimize to obtain the current topology of the NG network and the optimal parameters of the classifier; The combined loss function simultaneously considers the Elastic Weight Consolidation (EWC) loss, Dynamic Topology Preservation loss, Prototype Contrast Loss, and Multi-Class Cross-Entropy Loss. The loss function is as follows:
[0027] in, Indicates elastic weight consolidation loss, Represents dynamic topology preservation loss, Indicates prototype contrast loss, This represents the cross-entropy loss for multi-class classification. These are hyperparameters used to balance the weights of different components; Elastic weight consolidation loss for:
[0028] in, These are the parameters of the current classifier. In response to historical characteristics The optimal parameters, These are the diagonal elements of the Fisher information matrix, used to measure the importance of parameters. The regularization strength is used to control the degree to which old knowledge is retained; Dynamic topology preservation loss for:
[0029] The above formula indicates that in NG dynamic expansion and At the same time, maintaining the consistency of the topological structure between nodes ensures that neighboring nodes remain close in the feature space. Let the age of the edge be , The age threshold, Let the set of edges between nodes be . and For nodes and nodes The center of gravity; Prototype contrast loss for
[0030] in As input features, To and The center of gravity of nodes of the same type The set of centroids of all nodes. It is the Euclidean distance function.
[0031] Furthermore, when i When =1, construct the classification loss. Using the Adam algorithm to Optimize the classifier to obtain its optimal parameters;
[0032] This represents the total number of samples in the training set. Total number of pilot categories; The true label is one-hot encoded, if the sample Category The value is 1 if it is 1, otherwise it is 0. Predict samples for the model Category The probability is used as the output of softmax.
[0033] Furthermore, utilizing current data Extracted features Calculate features Distance to each NG node The input is a classifier, which uses the softmax activation function to predict the probability of belonging to a class. The output vector is:
[0034] The softmax activation function is used. These are the parameters of the classifier. A vector containing the probability of each class , Find the maximum value for the total number of items in the current category. Then this set of data Belongs to the Pilot-type.
[0035] Beneficial effects: 1. This method provides a small-sample class incremental learning approach that can achieve continuous recognition of new types of pilots by adding a small number of new class samples.
[0036] 2. In the feature extraction stage, spatial situation features and their statistical features within the time window of effective flight data are extracted based on prior knowledge, generating a set of spatial situation features for different pilots, which enhances the interpretability of feature selection.
[0037] 3. A dynamic topology adaptation mechanism is designed. When the distance between a new class sample and an existing prototype is greater than a threshold, new neurons are added, which improves the adaptability to the new class.
[0038] 4. A combined loss function was constructed, which weights and combines Elastic Weight Consolidation (EWC) loss, Dynamic Topology Preservation loss, Prototype Contrast loss and Classification loss. This can effectively avoid overfitting in small samples and forgetting old knowledge, and has good application prospects in pilot type recognition scenarios where data is scarce and needs to be continuously updated. Attached Figure Description
[0039] Figure 1 This is a flowchart of the present invention.
[0040] Figure 2 This provides the task scenario and the first-person perspective of the participants for Tacview playback.
[0041] Figure 3 The loss curves represent the training processes of the basic category model and the incremental learning model for new categories.
[0042] Figure 4 The basic category of GNG network topology.
[0043] Figure 5 This represents the incremental expansion results of the GNG network topology.
[0044] Figure 6 The decision boundaries for the three pilot types after incremental learning. Detailed Implementation
[0045] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.
[0046] In this application, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for the purpose of better describing this application and its embodiments, and are not intended to limit the indicated device, element, or component to having a specific orientation, or to be constructed and operated in a specific orientation.
[0047] Furthermore, the terms "installation," "setup," "equipped with," "connection," "linking," and "socketing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral structure; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of these terms in this application based on the specific circumstances.
[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0049] A continuous pilot identification method based on an incremental neural framework with topological information is proposed. First, in the feature extraction stage, spatial situation features and their statistical characteristics within a time window are extracted from valid flight data based on prior knowledge, generating spatial situation feature sets for different pilots. Second, in the topological network layer, dynamic clustering generates pilot behavior prototypes, and neuron distribution is adjusted through competitive learning; a dynamic topological adaptation mechanism is designed, adding neurons when the distance between a new category sample and an existing prototype exceeds a threshold. Third, elastic weight consolidation (EWC) loss and dynamic topological preservation loss are introduced to balance learning new categories and retaining old categories, avoiding catastrophic forgetting; prototype contrast loss is introduced to enhance the clustering of similar samples, widening the distance between dissimilar samples and avoiding overfitting caused by small sample sizes.
[0050] A persistent pilot type recognition method based on an incremental neural framework with topological information includes the following steps, and the flowchart of the method of the present invention is shown below. Figure 1 As shown: S1, Obtain the pilot category labeling dataset The ACMI format data was exported to Excel format using Tacview software at a sampling frequency of 20 Hz. The exported parameters included flight status parameters such as longitude, latitude, and altitude of both friendly and enemy aircraft, as well as operational status parameters such as roll and pitch angles, totaling 61 dimensions. Taking the pilot's evasion of an enemy missile as the decision point, data was extracted 2.5 seconds before and after that moment, and the pilot type of this data set was labeled as pilot category labeling data. Therefore, the data time window was 5 seconds. S2, from the current data Extracting relative spatial situational features, including relative altitude, relative distance, azimuth, elevation, and proximity rate. From current data... Extract local features, including energy characteristics, turning radius, and angular velocity. Calculate the statistical characteristics of all these features within a time window.
[0051] (1) The relative height is calculated using the following formula:
[0052] This refers to the altitude of the machine. For the first The altitude of the enemy aircraft For this machine and the first The altitude difference between the enemy aircraft.
[0053] (2) The relative distance is calculated using the following formula:
[0054] These are the local machine's geocentric rectangular coordinates. For the first The geocentric rectangular coordinates of the enemy aircraft For this machine and the first Spatial distance of enemy aircraft.
[0055] (3) The azimuth angle is calculated using the following formula:
[0056] Where E and N are obtained by transforming relative vectors to the local coordinate system (ENU coordinate system) of the machine:
[0057] in , indicating the first The enemy aircraft's coordinates relative to this aircraft. East (E) is along the latitude circle eastward, North (N) is along the longitude line northward, and Sky (U) is perpendicular to the ground upward. Longitude Latitude.
[0058] (4) The pitch angle is calculated using the following formula:
[0059] (5) The proximity rate is calculated using the following formula:
[0060] in, , , Let E be the east component of the velocity of the k-th enemy aircraft. The E component represents the speed of my machine. Let N be the north (N) component of the velocity of the k-th enemy aircraft. The north (N) component represents the speed of my machine.
[0061] From current data Extracting local features, including energy characteristics, turning radius, and angular velocity, (1) The formula for calculating energy characteristics is as follows:
[0062] in For the machine's speed, This refers to the altitude of the machine. This is the acceleration due to gravity.
[0063] (2) The turning radius is denoted as R, which can be directly obtained from the dataset. Read from the middle.
[0064] (3) The formula for calculating angular velocity is as follows:
[0065] Using a 5-second time window and a 20Hz sampling frequency, a total of 100 sampling points are used. The statistical characteristics of all the above features within the time window are calculated. The formula for calculating the mean is: The formula for calculating the standard deviation is: The formula for calculating skewness is: The formula for calculating kurtosis is: ;in This represents a specific feature at a given sampling point. This represents the mean of a certain feature within a time window. This represents the standard deviation of a certain feature within a time window. This indicates the skewness of a feature within a time window. This represents the kurtosis of a certain feature within a time window, where n is the number of samples within a time window; in this example, n=100. m The value of is a positive integer from 1 to 100.
[0066] S3, when i When =1, enter S4, when i If the value is greater than 1, proceed to step S5.
[0067] S4, construct a... Neural Gas (NG) network with 10 nodes , It describes the network topology. ( ) represents a node in the network. Let be the set of edges, representing the adjacency relationships between nodes. Calculate each feature. Distance to each node ,
[0068] in For each node's centroid, sort the distances from it in ascending order to obtain the following arrangement of nodes:
[0069] So, for each node Its center of gravity Updated to :
[0070] in For learning rate, The decay function, The updated centroid is a constant. Closer to features .
[0071] Features nearest node and the second nearest node The node's "age" and edge connections are updated. The "age" and the updates to the connected edges are as follows: ,
[0072] node The "age" and the updates to the connected edges are as follows: ,
[0073] and The connection between them is made by letting and corresponding age To create or maintain. If the neuron is... or The "age" is reset to 0; for other neurons, the age is increased by 1, and neurons with higher ages are deleted. For other edges, if... Exceeding the maximum age , The connection was deleted.
[0074] Through the above operations, each feature Assign to the nearest node, and allocate the feature space. Divide the area into unconnected regions and proceed to step S7.
[0075] S5, wheni When the threshold is greater than 1, design an adaptive threshold:
[0076] in , Historical characteristics The average error and standard deviation of the distribution of historical nodes.
[0077] Input current feature Find the nearest node Its center of gravity is Detect whether the current feature is significantly different from the distribution of existing nodes. If:
[0078] Then add a new neuron and proceed to step S6. If:
[0079] Proceed to step S7.
[0080] S6, Design the combination loss function Using the Adam algorithm
[15] to... Optimization is performed to obtain the current topology of the NG network and the optimal parameters of the classifier.
[0081] The combined loss function considers Elastic Weight Consolidation (EWC) loss, Dynamic Topology Preservation loss, Prototype Contrast Loss, and Multi-Class Cross-Entropy Loss simultaneously. The loss function is as follows:
[0082] in, Indicates elastic weight consolidation loss, Represents dynamic topology preservation loss, Indicates prototype contrast loss, This represents the cross-entropy loss for multi-class classification. These are hyperparameters used to balance the weights of different components. Elastic weight consolidation loss for:
[0083] in, These are the parameters of the current classifier. In response to historical characteristics The optimal parameters, These are the diagonal elements of the Fisher information matrix, used to measure the importance of parameters. The regularization strength is used to control the degree to which old knowledge is retained.
[0084] Dynamic topology preservation loss for:
[0085] The above formula indicates that in NG dynamic expansion and At the same time, maintaining the consistency of the topological structure between nodes ensures that neighboring nodes remain close in the feature space. Let the age of the edge be , The age threshold, Let the set of edges between nodes be . and For nodes and nodes The center of gravity.
[0086] Prototype contrast loss for
[0087] in As input features, To and The center of gravity of nodes of the same type The set of centroids of all nodes. It is the Euclidean distance function.
[0088] S7, Constructing the classification loss Using the Adam algorithm to Optimize the classifier to obtain its optimal parameters.
[0089]
[0090] This represents the total number of samples in the training set. This represents the total number of pilot categories. The true label is one-hot encoded, if the sample Category If it is 1, then it is 1; otherwise, it is 0. Predict samples for the model Category The probability is the output of softmax.
[0091] S8, using current data Extracted features Calculate features Distance to each NG node The input is a classifier, which uses the softmax activation function to predict the probability of belonging to a class. The output vector is:
[0092] The softmax activation function is used. These are the parameters of the classifier. A vector containing the probability of each class , Find the maximum value for the total number of items in the current category. Then this set of data Belongs to the Pilot-type.
[0093] Example: Based on the DCS simulation platform, three professionals with advanced air combat skills participated in the data collection, conducting 22 2V2 air combat missions. The mission scenarios and participants' first-person perspectives were replayed using Tacview. Figure 2 As shown.
[0094] The ACMI format data was exported to Excel format using Tacview software at a sampling frequency of 20 Hz. The exported parameters included flight status parameters such as longitude, latitude, and altitude for both friendly and enemy aircraft, as well as operational status parameters such as roll and pitch angles, totaling 61 dimensions. Data from the five seconds preceding the first attack by two participants was extracted. Features were calculated based on the S2 algorithm, and the topology network and classifier were initialized. Then, a third participant was used as a new class sample as input, and the topology network structure and classifier parameters were incrementally updated to test the model's generalization ability to the new sample, verifying the effectiveness of the incremental neural framework for topological information.
[0095] Using a uniform sampling method for different categories, 70% of the samples were selected as the training set and 30% as the test set. The accuracy of the basic model in identifying the old category on the test set, the accuracy of the incrementally learned model in identifying the old category, the accuracy in identifying the new category, and the accuracy in identifying the overall category were calculated.
[0096]
[0097] in The sum of all class samples in the test set. It is the total number of categories. To be correctly predicted as a category The number of samples.
[0098] Figure 3 The diagram shows the loss curves for training the basic category model and the incremental learning model for the new category. The GNG network topology for the basic category is shown below. Figure 4 As shown. When the input data differs significantly from the existing node distribution, it indicates the addition of a new pilot category. S4 is the judgment condition. In this case, new nodes need to be added to dynamically expand the NG network. The incremental expansion results of the GNG network topology are shown below. Figure 5 As shown, blue represents existing nodes, and green represents newly added nodes. The decision boundaries for the three categories after incremental learning are as follows: Figure 6 As shown, the different shapes represent the categories to which the pilots belong.
[0099] The accuracy rates of category prediction using the proposed TINet-PII method and the classification layer retraining method are shown in Table 1. As can be seen from Table 1, compared with the classification layer retraining method, the proposed TINet-PII method achieves a high recognition accuracy of 90.23% for new categories while maintaining the retention of old categories (recognition accuracy for old categories is 76.67%). The recognition accuracy for all categories reaches 82.92%, indicating that this method effectively avoids overfitting to small samples and forgetting of old knowledge.
[0100] Table 1. Class prediction accuracy of TINet-PII and retrained classification layer
[0101] The above description is merely a specific embodiment of the present invention, providing a detailed description of the invention. Parts not covered herein are conventional techniques. However, the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for persistent pilot type recognition based on an incremental neural framework with topological information, characterized in that, include: S1, obtain t Group pilot category labeled dataset , ; S2, from the current data Extracting relative spatial situation features and local features, and calculating the statistical features of all the above features within a time window, the resulting feature space is denoted as […]. ; S3, when i When =1, enter S4, when i If the value is greater than 1, proceed to step S5; S4, construct a... Neural gas network of nodes , It describes the network topology. ( ) represents a node in the network. Let the set of edges represent the adjacency relationships between nodes; calculate each feature. Distance to each node Each feature Assign to the nearest node, and allocate the feature space. Divide the regions into unconnected areas and proceed to step S7; S5, Design an adaptive threshold to detect the current feature. If the distribution of nodes is significantly different from the existing node distribution, and the difference is greater than the threshold, then a new category has emerged, and new nodes need to be added to dynamically expand the NG network, proceeding to step S6; if the difference is less than the threshold, proceed to step S7. S6, Design the combination loss function Using the Adam algorithm to Optimize to obtain the current topology of the NG network and the optimal parameters of the classifier, then proceed to step S8; S7, Constructing the classification loss Using the Adam algorithm to Optimize the classifier to obtain its optimal parameters; S8, using current data Extracted features Calculate features Distance to each NG node Input the data into the classifier to obtain the pilot's category.
2. The pilot type persistent recognition method based on a topological information incremental neural framework according to claim 1, characterized in that, Pilot Category Labeling Dataset The data was exported from ACMI format to Excel format using Tacview software at a sampling frequency of 20 Hz. The exported parameters included flight status parameters such as longitude, latitude, and altitude of both sides, as well as operational status parameters such as roll and pitch angles, totaling 61 dimensions. Taking the pilot's evasion of enemy missiles as the decision point, data was extracted 2.5 seconds before and after that moment, and the pilot type of this set of data was labeled as pilot category label data. Therefore, the data time window was 5 seconds.
3. The pilot type persistent recognition method based on a topological information incremental neural framework according to claim 2, characterized in that, From current data Extract relative spatial situation features, including relative altitude, relative distance, azimuth, elevation angle, and proximity rate; (1) The relative height is calculated using the following formula: This refers to the altitude of the machine. For the first The altitude of the enemy aircraft For this machine and the first The altitude difference between the enemy aircraft; (2) The relative distance is calculated using the following formula: These are the local machine's geocentric rectangular coordinates. For the first The geocentric rectangular coordinates of the enemy aircraft For this machine and the first The spatial distance between enemy aircraft; (3) The azimuth angle is calculated using the following formula: Where E and N are obtained by transforming relative vectors to the local coordinate system of the machine: in , indicating the first The enemy aircraft's coordinates relative to this aircraft; East E extends eastward along the latitude circle, North N extends northward along the longitude line, and U is perpendicular to the ground upwards. Longitude Latitude; (4) The pitch angle is calculated using the following formula: (5) The proximity rate is calculated using the following formula: in, , , Let E be the eastward component of the velocity of the k-th enemy aircraft. The E component of my machine's speed; Let N be the north component of the velocity of the k-th enemy aircraft. The N component represents the speed of my machine.
4. The pilot type persistent recognition method based on an incremental neural framework with topological information according to claim 3, characterized in that, From current data Extracting native features, including energy characteristics, turning radius, and angular velocity: (1) The formula for calculating energy characteristics is as follows: in For the machine's speed, This refers to the altitude of the machine. It is the acceleration due to gravity; (2) The turning radius is denoted as R, which can be directly obtained from the dataset. Read from; (3) The formula for calculating angular velocity is as follows: 。 5. The pilot type persistent recognition method based on a topological information incremental neural framework according to claim 4, characterized in that, Calculate the statistical characteristics of all the above features within the time window, including: A total of 100 sampling points were collected, with a time window of 5 seconds and a sampling frequency of 20 Hz. Calculate the statistical characteristics of all the above features within the time window: The formula for calculating the mean is: ; The formula for calculating standard deviation is: ; The formula for calculating skewness is: ; The formula for calculating kurtosis is: ; in This represents a specific feature at a given sampling point. This represents the mean of a certain feature within a time window. This represents the standard deviation of a certain feature within a time window. This indicates the skewness of a feature within a time window. This represents the kurtosis of a feature within a time window, where n is the number of samples within a time window.
6. The pilot type persistent recognition method based on a topological information incremental neural framework according to claim 5, characterized in that, when i When =1, construct a... Neural gas network of nodes , It describes the network topology. ( ) represents a node in the network. Let the set of edges represent the adjacency relationships between nodes; calculate each feature. Distance to each node , in For each node's centroid, sort the distances from it in ascending order to obtain the following arrangement of nodes: So, for each node Its center of gravity Updated to : in For learning rate, The decay function, For a constant, the updated centroid Closer to features ; Features nearest node and the second nearest node Update the "age" and edge information of the node; The "age" and the updates to the connected edges are as follows: ， node The "age" and the updates to the connected edges are as follows: ， and The connection between them is made by letting and corresponding age To create or maintain; if the neuron is or The "age" is reset to 0; for other neurons, the age is increased by 1, and neurons with higher ages are deleted; for other edges, if... Exceeding the maximum age , The edge was deleted. Through the above operations, each feature Assign to the nearest node, and allocate the feature space. Divide into unconnected regions.
7. The pilot type persistent recognition method based on a topological information incremental neural framework according to claim 6, characterized in that, when i When the threshold is greater than 1, design an adaptive threshold: in , Historical characteristics The average error and standard deviation of the distribution of historical nodes; Input current feature Find the nearest node Its center of gravity is Detect whether the current feature is significantly different from the distribution of existing nodes. If: Then add a new neuron.
8. The pilot type persistent recognition method based on a topological information incremental neural framework according to claim 7, characterized in that, when i When the value is greater than 1, design the combination loss function. Using the Adam algorithm [15] to... Optimize to obtain the current topology of the NG network and the optimal parameters of the classifier; The combined loss function simultaneously considers the Elastic Weight Consolidation (EWC) loss, Dynamic Topology Preservation loss, Prototype Contrast Loss, and Multi-Class Cross-Entropy Loss. The loss function is as follows: in, Indicates elastic weight consolidation loss, Represents dynamic topology preservation loss, Indicates prototype contrast loss, This represents the cross-entropy loss for multi-class classification. These are hyperparameters used to balance the weights of different components; Elastic weight consolidation loss for: in, These are the parameters of the current classifier. In response to historical characteristics The optimal parameters, These are the diagonal elements of the Fisher information matrix, used to measure the importance of parameters. The regularization strength is used to control the degree to which old knowledge is retained; Dynamic topology preservation loss for: The above formula indicates that in NG dynamic expansion and At the same time, maintaining the consistency of the topological structure between nodes ensures that neighboring nodes remain close in the feature space. Let the age of the edge be , The age threshold, Let the set of edges between nodes be . and For nodes and nodes The center of gravity; Prototype contrast loss for in For input features, To and The center of gravity of nodes of the same type The set of centroids of all nodes. It is the Euclidean distance function.
9. The pilot type persistent recognition method based on a topological information incremental neural framework according to claim 8, characterized in that, when i When =1, construct the classification loss. Using the Adam algorithm to Optimize the classifier to obtain its optimal parameters; The total number of samples in the training set. Total number of pilot categories; The true label is one-hot encoded, if the sample Category The value is 1 if it is 1, otherwise it is 0. Predict samples for the model Category The probability is used as the output of softmax.
10. The pilot type persistent recognition method based on a topological information incremental neural framework according to claim 9, characterized in that, Using current data Extracted features Calculate features Distance to each NG node The input is a classifier, which uses the softmax activation function to predict the probability of belonging to a class. The output vector is: The softmax activation function is used. These are the parameters of the classifier. A vector containing the probability of each class , Find the maximum value for the total number of items in the current category. Then this set of data Belongs to the Pilot-type.