Flight mode intelligent identification method based on IFPRM-SBLFS deep learning

Abstract

The invention relates to an intelligent flight mode recognition method based on IFPRM-SBLFS deep learning, solves the problem of flight state recognition from the perspective of time detection for the first time, and provides an intelligent flight mode recognition method combining sensitive boundary recognition and long flight sequence processing. According to the method, the limitation of a traditional method in single-mode recognition and composite maneuvering recognition is successfully broken, the tense boundary of the flight mode can be accurately positioned, online automatic recognition can be achieved, and the practicability and operation convenience of the model are greatly improved. The self-adaptive graph embedding module designed by the invention is mainly used for mining the spatial relationship of various flight parameters in the flight data, and can dynamically adjust the graph structure and adapt to different flight modes. Besides, a multi-scale transformer detector is introduced, so that the model has a multi-scale characteristic, abnormal modes during flight can be detected under different time scales, and fine flight characteristic changes can be captured.

Description

Technical Field

[0001] The present invention relates to the field of flight mode intelligent recognition methods, and in particular to a flight mode intelligent recognition method based on IFPRM-SBLFS deep learning. Background Art

[0002] Aircraft flight state recognition is to identify the flight state of the aircraft at any time during the flight through the flight parameters collected by the flight data recorder (FDR), and determine the starting and ending points of each state. This technology is crucial for modern aircraft and is widely used in fields such as over-limit monitoring, health monitoring, flight training, and usage monitoring. With the increase in demand, flight state recognition methods face huge challenges in terms of recognition accuracy, pattern boundary accuracy, recognition speed, and altitude.

[0003] Data-driven machine learning has been widely used in flight mode recognition, flight phase identification, flight trajectory prediction, anomaly detection, fault diagnosis, and pilot fatigue detection.

[0004] An accurate flight mode recognition method can assign a damage coefficient to each component according to the severity of the identified flight state, thereby more accurately estimating the actual state of fatigue use. Since the late 1990s, a variety of flight mode recognition methods have been proposed in the literature, which can be roughly divided into knowledge-based methods, threshold-based methods, and machine learning-based methods. Among them, the knowledge-based method, Lee et al., created and combined a universal maneuver library for fixed-wing aircraft to implement tactical maneuvers. They established this universal maneuver library in response to the various tactical maneuver characteristics required on the battlefield. In the creation of the maneuver library and the generation of modal input trajectories or flight paths, seventh-order polynomials were used for modeling and analysis. The research team conducted flight simulations of various maneuvers such as longitudinal circular flight, turning, and climb-turn-descent to verify that the combination of the constructed maneuver library can actually achieve tactical maneuvers. This research not only helps to promote the development of aircraft tactical maneuver flight technology, but also provides a theoretical basis for the revision of the Air Force Operations Manual. The threshold-based method combines the analysis of the collected data with prior domain knowledge to determine a set of thresholds suitable for the intended purpose. These methods have overlapping standards when distinguishing two or more systems, which can lead to improper identification, and are difficult to expand as the number of rules required for advanced identification continues to increase. The machine learning-based method automatically identifies the basic patterns in the signal by training the classifier, thereby improving the flight state recognition ability. Jiao et al. proposed a flight attitude recognition method based on audio features. Using the newly recorded drone attitude audio dataset, a lightweight convolutional neural network with an integrated attention mechanism was established for flight attitude recognition, with an accuracy of 98.81%. In view of the fact that the current maneuvering action recognition method is limited by the existing maneuvering action division criteria, Yin Zhiping et al. proposed an automatic maneuvering action division and recognition method based on the time series based on the fusion of flight trajectory projections in the horizontal and plumb planes. This maneuvering action division and recognition method can improve the recognition efficiency of flight actions while ensuring the division efficiency. Zhang Yulu et al. proposed a recognition method based on random forests for the problem of flight state recognition of the gliding phase of near-space hypersonic reentry gliders. This method has high recognition accuracy, but when there are errors in the estimation of motion parameters, the recognition accuracy will decrease to a certain extent.

[0005] Existing research methods have deficiencies in six aspects, including knowledge dependence, generalization ability, pattern continuity, complex patterns, and pattern boundaries. Moreover, with single-point or equal-length window flight parameters as input, the complexity of the criterion increases exponentially, leading to the risk of overfitting. Machine learning-based methods mainly rely on classification frameworks to achieve point-to-mechanism or multi-point-to-single-mechanism predictions, ignoring the continuity and decision boundaries of the system, and are difficult to deal with complex systems. Summary of the invention

[0006] In order to effectively solve the above problems, the present invention provides a flight mode intelligent recognition method based on IFPRM-SBLFS deep learning.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] An intelligent flight mode recognition method based on IFPRM-SBLFS deep learning includes the following steps:

[0009] Step 1: collect flight data, and manually annotate the flight mode of each time stage in each flight data to obtain training samples, and multiple training samples constitute a training set;

[0010] Establish an IFPRM-SBLFS deep learning-based flight mode intelligent recognition model including an adaptive graph embedding model, a self-attention transformer module, a denoising deep multi-scale autoencoder, a boundary regressor, and a pattern classifier;

[0011] Step 2: Input the training samples into the adaptive graph embedding model, extract the short-term and long-term flight parameter change trends respectively through sliding windows of different lengths, use spectral graph convolution to extract spatial features of different time scales and perform multi-scale feature splicing to obtain the comprehensive spatial features at each moment;

[0012] Step 3: Input the comprehensive spatial features into the self-attention transformer module, analyze the short-term and long-term temporal dependencies through the self-attention transformer module and the denoising deep multi-scale autoencoder, and generate multi-scale features;

[0013] Step 4: Input the multi-scale features into the boundary regressor and the mode classifier respectively, the mode classifier outputs the flight mode to which each time stage in the flight data belongs, and the boundary regressor outputs the boundary of the flight mode to which each time stage in the flight data belongs;

[0014] Step 5: Compare the output results of the boundary regressor and the pattern classifier with the manual annotation results, calculate the classification weighted focal loss of the boundary regressor and the joint spatiotemporal intersection loss of the pattern classifier, and adjust the parameters of the boundary regressor and the pattern classifier through back propagation;

[0015] Step 6: Repeat steps 1-5 to train the boundary regressor and pattern classifier by minimizing the loss function to obtain the trained boundary regressor and pattern classifier, and finally obtain the trained flight mode intelligent recognition model based on IFPRM-SBLFS deep learning;

[0016] Step 7: Input the flight data of the flight to be identified into the trained flight mode intelligent recognition model based on IFPRM-SBLFS deep learning, and the model outputs the flight mode belonging to each time stage in the flight.

[0017] Furthermore, in step 1, the flight modes include 11 modes: take-off, level flight, climb, climb and rotation, sledgehammer, 360° roll, barrel turn, climb and roll, half roll, somersault, lazy eight, landing and rotation, dive, and landing.

[0018] Furthermore, each flight data includes the real-time altitude, position, speed, roll angle and pitch angle data of the aircraft corresponding to all moments in a flight.

[0019] Furthermore, a multi-scale sliding window mechanism is adopted in the adaptive graph embedding model to extract the short-term and long-term flight parameter change trends respectively through sliding windows of different lengths, and integrate the short-term and long-term flight parameter change trends into the graph construction process; in the original adjacency matrix construction process, the Gaussian kernel function is introduced into the Euclidean distance to construct the adjacency matrix, thereby enhancing the description of nonlinear relationships, and the adjacency matrix is ​​adaptively weighted in combination with the attention mechanism.

[0020] Furthermore, in the self-attention transformer module, a transformer is used to replace the traditional short window filter. The self-attention transformer module takes the comprehensive spatial features as input, and weights and averages all the features at each moment to generate new features at each moment. Specifically, the weight of each feature during weighting is the correlation between the current feature and other features.

[0021] Furthermore, the denoising deep multi-scale autoencoder includes an encoder, a decoder and a hidden layer, which are used to encode and decode the feature data output by the self-attention transformer module to obtain the potential features of the data, and output the corresponding reconstructed data through the decoder to generate multi-scale features.

[0022] Furthermore, the network structure of the pattern classifier is expanded from the traditional three-layer one-dimensional convolutional network to a four-layer structure, and a multi-scale feature fusion module is introduced in each layer to enhance the network's ability to extract contextual information from different scales. By adopting a variety of convolution kernel sizes, the network can simultaneously capture local and global features, and the output of each layer is enhanced by a multi-head attention mechanism. Eight attention heads are used to perform fine-grained allocation of feature weights, further improving the feature expression capability of the classifier.

[0023] Furthermore, the boundary regressor adopts a four-layer network structure, each layer is processed using a multi-scale fusion module with convolution kernel sizes of 3, 5, and 7, and a context-based range constraint module is introduced in each layer to ensure that the boundary dynamic information in the features is effectively captured and processed, and the output of each layer is enhanced by a multi-head attention mechanism. Through the feature enhancement of eight heads, the features at different scales can be fully expressed. In each layer of the network, in addition to the traditional ReLU activation function, Leaky ReLU is also introduced to deal with the gradient disappearance problem. The output range of the boundary regressor is adaptively adjusted according to the training data, where the first layer maintains the initial range and the subsequent layers are dynamically expanded according to the data distribution.

[0024] After adopting the above technical solution, the present invention has the following advantages compared with the prior art:

[0025] (1) The present invention solves the problem of flight state recognition from the perspective of time detection for the first time. The present invention proposes an intelligent flight mode recognition method that combines sensitive boundary recognition and long flight sequence processing, which successfully overcomes the limitations of traditional methods in single mode recognition and complex maneuver recognition, and can accurately locate the temporal boundaries of flight modes.

[0026] (2) Aiming at the uniqueness of flight data, the present invention designs the IFPRM-SBLFS model. The model has a high degree of automation and intelligence, avoids dependence on expert knowledge and experience, enables online automatic recognition on the aircraft, and greatly improves the practicality and ease of operation of the model.

[0027] (3) The adaptive graph embedding module designed by the present invention is mainly used to mine the spatial relationship of multiple flight parameters in flight data, and can dynamically adjust the graph structure and adapt to different flight modes. In addition, a multi-scale transformer detector is introduced to make the model multi-scale, which can detect abnormal patterns during flight at different time scales and capture subtle changes in flight characteristics.

[0028] (4) The proposed model adopts classification weighted focal loss, which effectively reduces the error caused by category imbalance in the flight mode classification process. At the same time, in order to improve the positioning accuracy of the flight mode boundary, a spatiotemporal intersection loss function is designed, which combines the intersection of spatiotemporal information to reduce the flight mode loss, making the recognition result more accurate and robust.

[0029] (5) Classification weighted focal loss increases the weight of difficult-to-classify samples, making the model pay more attention to a small number of important samples, thereby reducing its dependence on easy-to-classify samples and avoiding overfitting; joint spatiotemporal intersection loss and adaptive graph embedding improve the model's ability to understand complex features by capturing long-sequence spatiotemporal dependencies, enabling it to maintain good generalization at different time and space scales; denoising deep multi-scale autoencoder uses a denoising mechanism to strengthen the model's capture of essential features and filter out noise interference, thereby further improving the model's robustness and generalization ability and preventing overfitting.

[0030] (6) Based on a real flight dataset, the boundaries of the flight states were marked, and the significant advantages of the IFPRM-SBLFS model in terms of precision and accuracy were verified through experiments, fully demonstrating the superiority and broad application prospects of the proposed method in complex flight state identification.

[0031] The present invention is described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Schematic diagram of the center of gravity trajectory of an aircraft during a flight.

[0033] Figure 2 is a flow chart of the method of the present invention;

[0034] Figure 3 It is a visualization result diagram of different pattern recognition; DETAILED DESCRIPTION

[0035] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0036] 1 Flight Dataset Description

[0037] The implementation of IFPRM-SBLFS deep neural network relies on a large amount of labeled data and model training. Flight data is real-time flight parameters collected from hundreds of flights. In order to accurately obtain the real-time status of the aircraft. In order to restore the three-dimensional flight trajectory and attitude of the aircraft, we manually annotate each moment in the flight trajectory according to the state definition and feature description in the flight training manual, and finally form a standardized annotated data set.

[0038] 1.1 Flight trajectory

[0039] The flight trajectory of the aircraft can be accurately reproduced by simulating the aircraft's real-time altitude, north latitude, east longitude, speed, azimuth, pitch angle, yaw angle, etc. in three-dimensional space. Finally, combined with the three-dimensional trajectory and key parameters of each sortie, the flight status contained in each sortie is annotated one by one.

[0040] 1.2 Dataset Description

[0041] Due to the diversity of usage scenarios and mission objectives of different aircraft, the flight states they contain are also different. Therefore, we defined 11 single flight modes based on the mission requirements in actual flight sorties. These modes include basic flight states such as takeoff, level flight, climb, roll, and landing. At the same time, some composite modes are also defined, such as climb and rotate, landing and rotate, and climb and roll. In addition, there are some modes called maneuvers, such as lazy eight, half roll inversion, and barrel turn. To avoid ambiguity, we refer to all the above defined actions as flight modes. Figure 1 The flight trajectory of a sortie is shown in Figure 1. From takeoff to landing, a total of 11 states are experienced, and each adjacent state is connected by a dotted line with an arrow.

[0042] It should be noted that these states do not cover all flight cycles of the entire flight sortie. For the unlabeled segments in the flight sortie, we mark them as “others”. Therefore, a total of 11 categories are included in the training and testing process of IFPRM-SBLFS.

[0043] 2 Flight mode recognition method based on IFPRM-SBLFS neural network

[0044] Figure 2 The implementation process of the IFPRM-SBLFS neural network in flight mode recognition is demonstrated. First, the adaptive graph embedding module segments the flight parameter sequence through a multi-scale sliding window to capture spatial features at different time scales. Then, the denoising deep multi-scale autoencoder module extracts the short-term and long-term time-dependent features of the flight sequence to form a multi-scale feature representation. Finally, these features are passed to the boundary regressor and mode classifier modules to output mode boundaries and classification results. The modules work closely together through feature sharing and weight adjustment to ensure accurate recognition of flight modes from multiple dimensions.

[0045] like Figure 2 As shown in the figure, the IFPRM-SBLFS model specifically collects multiple flight parameters through FDR, and after data preprocessing and annotation, obtains the C variable time series X. The process is divided into three stages: (1) Through the adaptive graph embedding technology, the subsequence of each moment is truncated using overlapping windows to extract spatial features, such as g:X→Zs, where Zs∈RCs×Ls represents the spatial feature vector, and Cs and Ls are the dimensions respectively; (2) The short-term and long-term temporal dependencies are analyzed through the denoising deep multi-scale autoencoder to generate multi-scale features, such as h:Z s →Z t =Z 1t ,Z2t ,...,Z Mt , where Z mt ∈RC t ×L m , M is the total number of scales, L m is the feature length of the mth layer; finally, the features of different layers are input into the boundary regressor and classifier to predict their boundaries, that is, d:Z t →Y. The feature weights are determined by calculating the loss and performing back propagation. (3) In the inference phase, the model outputs the type and boundary of the flight mode. To eliminate incorrect predictions, any bounding box with a prediction probability lower than a preset threshold will be deleted, and post-processing techniques will be used to eliminate incorrect predictions and reduce redundancy.

[0046] In practical applications, traditional feature extractors have difficulty capturing long-term time dependencies, large differences in flight cycle lengths, and the impact of category imbalance on model recognition. To address these issues, the present invention innovatively proposes methods such as adaptive graph embedding, multi-scale encoders, weighted focal loss, and joint spatiotemporal intersection loss, thereby achieving multi-scale feature fusion and effectively improving recognition accuracy and robustness under category imbalance conditions. These innovative methods not only improve the accuracy of flight mode recognition, but also significantly enhance the adaptability and stability of the model.

[0047] 2.1 Adaptive Graph Embedding Model

[0048] As different flight parameters change, the dependency structure of the graph is updated dynamically. Therefore, we introduced a dynamic weight allocation strategy based on the attention mechanism, so that the graph structure can be adjusted in real time as the data changes, thereby further refining the spatial dependencies between different parameters. A single-scale sliding window will ignore the spatial dependencies at different time granularities. To solve this problem, we adopted a multi-scale sliding window mechanism to extract short-term and long-term flight parameter change trends through sliding windows of different lengths, and integrated them into the graph construction process. In the original adjacency matrix construction process, the Euclidean distance is difficult to accurately reflect the complex parameter dependencies. We introduced the Gaussian kernel function to construct the adjacency matrix, thereby enhancing the description of nonlinear relationships, and combined the attention mechanism to adaptively weight the adjacency matrix. The improved adaptive graph embedding method is mainly based on the original method, further enhancing the adaptability and accuracy of the model, especially for capturing the complex dependencies between different flight parameters and optimizing the graph structure construction process. The following is the improved adaptive graph embedding model.

[0049] For the flight sequence X, a multi-scale sliding window (length d1, d2, ..., d k , step length is s), and multiple truncated segments of different scales are obtained Through the projection matrix Truncated segments of different scales are embedded to obtain latent Euclidean space representations.

[0050]

[0051] The Gaussian kernel function is used to calculate the nonlinear distance between different flight parameters for the embedded truncated segments:

[0052]

[0053] Here, σ is the bandwidth parameter, which is used to control the smoothness of the Gaussian kernel.

[0054] In order to further improve the accuracy of the adjacency matrix, the attention mechanism is introduced to adaptively weight the adjacency matrix to obtain the final adjacency matrix A:

[0055] A ij =α ij ·D ij (3)

[0056] Among them, α ij is the weight calculated by the self-attention mechanism, which indicates the strength of the association between parameter i and parameter j. ij The calculation is as follows:

[0057]

[0058] f(z i ,z j ) means z i and z j The scoring function for the similarity between .

[0059] Based on the improved adjacency matrix A, spectral graph convolution is used to extract spatial features of different scales. The graph convolution of each scale i is defined as:

[0060]

[0061] in, is the normalized Laplacian matrix, is the Chebyshev polynomial expansion, θ m are the corresponding graph convolution parameters.

[0062] Concatenate graph convolution features of all scales to form comprehensive multi-scale spatial features:

[0063] Z=concat(Z′1,Z′2,…,Z′ k ) (6)

[0064] Finally, the concatenated features are reduced in dimension through the fully connected layer to obtain the comprehensive spatial features h at each moment. t:

[0065] h t =σ(W f Z+b f ) (7)

[0066] Among them, W f is the weight of the fully connected layer, b f is the bias term and σ is the ReLU activation function.

[0067] 2.2 Flight parameters

[0068] The online spatiotemporal flight parameters detected by ground ship radar, infrared, electronic countermeasure and other tracking equipment include four categories: (1) Position parameters, which describe the GPS north latitude, east longitude, speed and altitude of the aircraft in three-dimensional space; (2) Attitude parameters, which describe the real-time attitude changes of the fuselage around three coordinate axes, heading angle, azimuth angle, pitch angle, roll angle, aileron angle, rudder angle, elevator angle; (3) Maneuvering parameters: describing the aircraft's yaw angle, pitch angle, roll angle, engine thrust, and fuel consumption rate; (4) Aviation engine parameters, engine speed, oil pressure, gasoline pressure, oil temperature, intake pressure, and carburetor temperature. There are a total of 22 original flight hyperparameters. At the same time, in order to improve the learning efficiency of the neural network, we uniformly standardize each parameter according to its range of variation so that the input parameter values ​​of all channels have the same order of magnitude. Before training the model, we introduced more flight data sets for training and testing, including the DJI Mavic data set and the UAV123 data set. These data sets not only expand the application scope of the model, but also provide a variety of flight scenarios to better verify the robustness of the model. In addition, we collect flight parameters in real time from the flight data recorder (FDR) and perform multi-dimensional verification on the flight data recorded by ground radar, optoelectronic equipment, electronic countermeasures and other tracking equipment. 70% of the flight sorties of all flight data sets are used as training data sets. train , and the remaining 30% is used as the test data set DATE test , while keeping the dataset flight mode distribution the same for both the training and test sets.

[0069] 2.3 Multi-scale Transform Detector

[0070] Capturing local spatial features of multivariate time series through adaptive graph embedding. The next step is to combine temporal relationships to obtain more discriminative features and use these features to detect flight status. There are currently two problems. First, due to the limited receptive area, traditional feature extractors cannot capture long-term temporal dependencies, so the detection performance for long periods is low. Second, in practical applications, the durations of different flight periods vary. Simple movements are usually slow and steady, while the flight patterns of aircraft are fast and diverse. Single-scale features cannot detect multiple durations simultaneously. To address these issues, the specific model design will be introduced below.

[0071] 2.4 Self-Attention Transformer Module

[0072] With the development of the self-attention mechanism, many studies have focused on exploring changes in the attention operator to further improve model performance. The Patchwise attention mechanism uses all feature information in the local area as attention weights, replacing the original softmax operation. The window attention adopted by Swin-Transformer enables the markers in the same local window to maintain the same receptive field, thereby saving computational costs and speeding up reasoning. Zhao and Nasiri considered global attention to preserve long-distance dependencies within a single layer. These improvements are effective under specific model architectures.

[0073] For the first problem, we replace the traditional short window filter with a transformer, which transforms the spatial feature Z s As input, the temporal relationship of all moments is calculated using the self-attention mechanism. Therefore, the model has the ability of long-term temporal interaction. Specifically, the feature of a certain moment can be calculated by the weighted average of the features of all moments, where the weight is the similarity between the target feature and other features.

[0074] Generally, given With C s channels and L s time steps, project it into the query vector Q, key vector K and value vector V, and learn the matrix as follows:

[0075]

[0076] Where D q =D k Then scale the self-attention The calculation formula is

[0077]

[0078] Convolution and self-attention are two powerful core techniques for representation learning. The first stage of both modules contains similar operation modes. More importantly, the computational complexity of the first stage (the square of the channel size) is dominant compared to the second stage. This observation naturally leads to a new idea of ​​combining these two techniques, that is, proposing a mixed model that enjoys the benefits of both self-Attention and Convolution (mixSAC) while minimizing the computational overhead compared to pure convolutional or self-attention counterparts. Our proposed (a mixed model that enjoys the benefit of both self-Attention and Convolution, mixSAC) is independent of the self-attention formula. Specifically, the attention weights can be summarized as:

[0079] (Small time period)

[0080]

[0081] (Fixed time period)

[0082]

[0083] (Global Information)

[0084]

[0085] Where [·] represents feature concatenation, φ(·) represents two linear projection layers with nonlinear activation in between, and W k (i, j) is the receptive field of each query token, and W represents the entire feature information.

[0086] In addition to the self-attention module, the transformer module also includes a layer normalization module, a fully connected module, and a residual connection module. The entire calculation process is as follows

[0087]

[0088] Where SA(·) represents the self-attention module, Norm(·) represents the layer normalization, and MLP(·) represents the fully connected module. in , Z' and Z out They are input features, median features, and output features respectively.

[0089] The input of the self-attention transformer module is a spatial feature matrix Z. The dimension of this matrix is ​​C×L, where C is the number of channels and L is the number of time steps. This input is usually obtained from the previous data processing stage and is used to represent the features obtained at a certain time step.

[0090] After being processed by the self-attention mechanism, the output is a new feature representation. The self-attention mechanism calculates the correlation (i.e., weight) between the input features and performs a weighted average on the input features to generate a contextual representation for each time step. Specifically, the weights are calculated based on the similarity between the input features, and then these weights are applied to the input feature data, and finally the newly generated feature Z' is output.

[0091] 2.5 Denoising Deep Multi-Scale Autoencoder

[0092] Although the transformer block can capture the global temporal dependencies in the flight sequence and solve the challenges of limited receptive field of traditional feature extractors (such as CNN), the resulting features are of a single scale and are difficult to meet the needs of multi-duration detection mechanisms. As a general characteristic of deep models, shallow features have higher resolution in short-term mechanisms but limited receptive fields in long-term mechanisms; in contrast, deep features have wider receptive fields in long-term mechanisms but lower resolution in short-term mechanisms. Therefore, the model can generate multi-scale features to better adapt to detection needs at different time scales. For a given flight data, the flight parameter expression matrix for each time period is denoted by X∈R V×G Represented as, where V is the flight parameter and G is the dimension of each parameter. The autoencoder consists of an encoder, a decoder, and a hidden layer, which is used to encode and decode the flight time series data to obtain the latent features of the data and output the corresponding reconstructed data through the decoder. Specifically, in the encoder, we input data X0 and get the output data Z in the encoding layer a , the calculation formula is as follows:

[0093] Z α =φ(W1X0+b1) (14)

[0094] Where φ is the LeakyReLU activation function, w2 is the weight matrix of the coding layer, and b2 is the bias of the coding layer. X0 = X + N, where N is uniformly distributed noise. Z a Mapped to Z through the hidden layer b , the formula is as follows:

[0095] Z b =φ(W2Z a +b2) (15)

[0096] Among them, φ is the LeakyReLU activation function, w2 is the weight matrix of the hidden layer, and b2 is the bias of the hidden layer.

[0097] Then, the output data Z of the same dimension as the encoding layer is obtained through the decoding layer c And reconstruct data X, the formula is as follows:

[0098] Z c =φ(W3Z b +b3) (16)

[0099] X'=φ(W4Z c +b4) (17)

[0100] In formulas (16) and (17), φ is the LeakyReLU activation function, w3 and w4 are the weight matrices of the decoding layer, and b3 and b4 are the biases of the decoding layer.

[0101] In order to improve the training effect of the autoencoder and effectively integrate the information between the encoding layer and the decoding layer, we use the following loss function to optimize the network:

[0102]

[0103] Where n represents the number of cells, X i represents the input feature of cell i, X′ i represents the reconstructed features of cell i, Z ai represents the features extracted by the encoding layer for cell i, Z ci Represents the features extracted by the decoding layer for cell i.

[0104] The output of the encoder produces multi-scale features, each representing a different level of features obtained by passing the data through multiple encoding layers.

[0105] In the encoder, short-layer features have higher resolution for short-term dependencies and are suitable for capturing rapidly changing features, while long-layer features have more advantages in long-term dependencies and can understand patterns with longer durations. This design enables the encoder to effectively capture multi-scale features and provide support for subsequent feature fusion and pattern recognition.

[0106] Through this design, the self-attention transformer module can effectively capture long-term dependencies and output important features, while the encoder outputs multi-scale features to enable the model to adapt to the needs of multiple temporal pattern recognition. The combination of the two greatly improves the performance of the model when recognizing complex flight patterns.

[0107] 2.6 Regressors and Classifiers

[0108] In the present invention, the network structure of the classifier is expanded from the traditional three-layer one-dimensional convolutional network to a four-layer structure, and a multi-scale feature fusion module is introduced in each layer to enhance the network's ability to extract rich contextual information from different scales. The design adopts a variety of convolution kernel sizes to enable the network to capture local and global features simultaneously, thereby significantly improving the accuracy of feature extraction. In addition, the output of each layer is enhanced by a multi-head attention mechanism, and eight attention heads are used to perform fine-grained allocation of feature weights to further enhance the feature expression ability of the classifier. In order to improve the stability of the network, the present invention adopts layer normalization and combines ELU activation function for nonlinear processing to reduce information loss, especially showing significant advantages when processing negative features. The final output channel number is set to 1 to represent the predicted probability of each time step. In addition, a parameter sharing strategy is introduced in the network structure to reduce the consumption of computing resources, and a GELU activation function is used in the last layer to ensure that the output probability range is limited to between 0 and 1, thereby improving the accuracy and stability of the results.

[0109] At the same time, the boundary regressor adopts a similar four-layer network structure, with each layer processed by a multi-scale fusion module with convolution kernel sizes of 3, 5, and 7, and a context-based range constraint module is introduced at each layer to ensure that the boundary dynamic information in the features is effectively captured and processed. The regressor also uses the multi-head attention mechanism consistent with the classifier, and through the feature enhancement of eight heads, the features at different scales can be fully expressed. In each layer of the network, in addition to the traditional ReLU activation function, Leaky ReLU is also introduced to effectively deal with the gradient disappearance problem, especially when dealing with negative outputs, which significantly improves the stability of training. To further optimize the processing of complex regression tasks, the output range of the boundary regressor is adaptively adjusted according to the training data, where the first layer maintains the initial range, and the subsequent layers are dynamically expanded according to the data distribution to ensure that the model effectively copes with prediction deviations under different input conditions, thereby improving the regression effect.

[0110] 2.7 Classification Weighted Focal Loss and Joint Spatial-Temporal Intersection Loss

[0111] In the deep learning framework, the last step is to calculate the loss function and optimize the model parameters by minimizing the loss. Since the model aims to achieve both pattern recognition and boundary localization at the same time, the loss consists of two parts: classification loss and regression loss.

[0112] In binary classification, class imbalance is a common challenge, especially when some classes appear much more frequently than other classes, which can cause the model to favor high-frequency classes and ignore low-frequency classes. To solve this problem, the present invention adopts Categorisation of Weighted Focal Point Losses (CWFPL). This loss function adjusts the weight of the loss so that the model pays more attention to difficult-to-classify samples and reduces the attention to easy-to-classify samples. However, the frequency of occurrence of different patterns shows an unbalanced long-tail feature, because the occurrence rate of simple patterns is higher than that of acrobatic patterns. In addition, the duration of different patterns is also different. Both of these situations will lead to class imbalance problems. To overcome this problem, K-way binary classification chooses Categorisation of Weighted Focal Point Losses.

[20] , which is widely used in objective detection to deal with unbalanced samples. The core of classification weighted focal loss lies in the adjustment of weight factors, which reduces the focus on majority class samples and makes the model pay more attention to "hard samples" and minority class samples.

[0113]

[0114] Where T represents the total number of points in all M layers; p ij ,y ij ∈{0,1} respectively represent the probability and label of determining whether point i belongs to the jth layer. α is used to reduce the weight of majority samples, and γ makes the model pay more attention to hard samples.

[0115] In addition to classification loss, the boundary regression task also requires an effective metric to evaluate the consistency between the predicted boundary and the true boundary. Temporal Intersection over Union (TIoU) originates from the IoU metric in image object detection and is used to measure the overlap rate between the predicted bounding box and the true bounding box. The purpose of TIoU is to evaluate the degree of overlap between the prediction mechanism and the ground truth label by measuring the prediction accuracy of the model in the temporal dimension. The calculation formula of TIoU is as follows:

[0116]

[0117] Among them, g j is an indicator function used to determine whether the sample contains the result hotspot number 1 (P) associated with the j pattern. j It represents the time point in the sample that is most likely to be predicted as mode j.

[0118] I(·) indicates the prediction result one-hot(P) j with the ground truth label y jThe intersection of . U(·) represents the prediction result No. 1 hotspot (P) j with the ground truth label y j of union.

[0119]

[0120] Among them, TPOS represents the positive point number, 1 i is an indicator function, which is equal to 1 when the value of point i is positive and equal to 0 when the value of point i is negative. is a regularization term used to bring the predicted center closer to the actual center, where d is the offset between the predicted center and the ground truth center, and c represents the joint length.

[0121] The total loss is the weighted sum of the classification loss and the regression loss. The weighted sum of the losses is shown in (23)

[0122] L=LCLS+wLREG (23)

[0123] Where w is a weighting coefficient used to balance the training process for better performance. Then, the model is trained by minimizing the overall loss. For ease of illustration, in the training algorithm of our proposed IFPRM-SBLFS, during the inference phase, the model also outputs a pattern instance for each point i in all M layers, where score i =max p i Represents the confidence score. As with anomaly detection, non-maximum suppression (NMS) is used for post-processing to retain the predictions with the highest confidence scores in the neighboring area and remove redundant predictions with high overlap. In addition, a confidence threshold is set for secondary screening to remove instances with low confidence scores and finally select several non-overlapping system boxes with high confidence.

[0124] 2.8 Evaluation Metrics

[0125] Unlike traditional methods, TRDN can predict either the system type or the system boundary. Therefore, the performance of system detection depends on these two aspects. A prediction is considered positive only when both the confidence score and the TIoU value of the prediction exceed the corresponding preset thresholds. On this basis, we will compare the prediction results with the ground truth, resulting in three cases: true positive, false positive, and false negative. Therefore, the calculation formulas for evaluation indicators such as precision, recall, and F1 score are:

[0126]

[0127] The above evaluation indicators are highly correlated with the preset confidence and tIoU critical values. The lower the threshold, the higher the recall rate and the lower the precision. In order to more comprehensively evaluate the performance of the model, we use the mean average precision (mAP), a commonly used evaluation indicator in objection detection. (mAP) is a commonly used evaluation indicator in objection detection. (mAP) is a commonly used evaluation indicator in objection detection. Specifically, all predictions of the same pattern type are arranged from high to low according to the confidence score. The dividing line divides the prediction into two parts, and the part with higher confidence participates in the calculation process. The precision and recall rates are calculated and plotted as a "PR graph". In the figure, all precision and recall rates form a curve, and the area under the curve is called AP. Finally, the calculation formula of mAP is:

[0128]

[0129] Where K is the number of pattern categories. i In the following study, we calculated the mAP value when the TIoU critical value was 0.5:0.1:0.9 to evaluate the performance of the model.

[0130] 3-mode extraction

[0131] 3.1 Probabilistic Identification

[0132] As a fragment analysis method, the prediction results of IFPRM-SBLFS are not independent at different time points. The prediction results of one time point will be affected by the adjacent time points. Therefore, in tasks that do not require high real-time performance, such as flight quality assessment, we can use a fine-grained overlapping sample generation strategy to select a large number of samples containing the same target time point and predict their results separately. Then, by fusing the prediction results of these samples, a more robust single-point prediction result can be obtained. By extending this strategy to the entire flight, the probability of all time points in the flight belonging to different flight states can be calculated as a preliminary recognition result.

[0133] 3.2 State Extraction

[0134] First, set a decision threshold p t , filter the fragments for each pattern. p tSet to 0.5. In order to ensure that all sub-segments extracted in this process have a high confidence coefficient, thereby providing a reliable basis for subsequent sub-segment splicing and expansion, we need to further process free points and system boundaries to improve the continuity and integrity of system segments. For free points that exist between the same pattern segments, if their duration is much shorter than the pattern duration, we will directly merge the two segments. For undefined intervals between different pattern segments, we can gradually expand their pattern boundaries by judging the probabilities of different patterns until the two segments are connected end to end. Finally, we can obtain all continuous and complete fine-grained pattern segments in the entire flight.

[0135] 4 Results Analysis

[0136] 4.1 Experimental setup

[0137] Table 1 lists the model structure and its hyperparameters in the present invention. For adaptive graph embedding, the length d of the sliding window, the step size s and the padding p are set to 11, 1 and 5 respectively. The order of the Chebyshev polynomial O of all graph convolutional layers is 4. The transformer encoder consists of 8 blocks, where the down-sampling rate r of the first four blocks is set to 1 and r of the remaining layers is set to 2. In addition, range constraints are applied in the last 8 blocks to obtain multi-scale features. Then, both the classifier and the regressor take it as input and use four convolutional layers (kernel size k=4, step size s=1, padding p=1) to predict the regime type and boundary. in and out represent the number of input and output channels.

[0138] The IFPRM-SBLFS model is based on Python 3.6, Pytorch 1.10, and CUDA 11.3. The experiments are conducted using a GeForceRTX 3070 GPU. During the training phase, the hyperparameters α, γ, and ω are empirically set to 0.25, 2, and 1, respectively.

[0139] Table 1 Model structure and hyperparameters

[0140]

[0141]

[0142] When the tIoU threshold and confidence threshold are set to 0.8 and 0.6 respectively, the precision, recall and F1 scores of various mechanisms are shown in Table 2. It can be seen that most flight modes achieve high precision and recall, indicating that the IFPRM-SBLFS neural network can detect almost all potential flight modes, and its detection results are accurate and reliable. Among them, the precision of all flight modes exceeds 97%, and the average result reaches 99.07%. The average result of recall for all flight modes is 91.74%. Considering the F1 score, the F1 score of all flight modes exceeds 94%, and the average result reaches 95.26%. The good results obtained on the test mechanisms of different aircraft cases and flight sorties show that the IFPRM-SBLFS neural network has good generalization ability. In addition, it is worth noting that the detection results can be adjusted by choosing different confidence thresholds. When FRR is used for flight training, the threshold should be set larger to obtain accurate detection results. When FRR is used for actual flight missions, the threshold should be lower to cover all actual states as much as possible.

[0143] Table 2 Comparative recognition results of IFPRM-SBLFS and other methods for single flight mode

[0144]

[0145] Based on the above five categories of indicators, we evaluate the performance of the proposed IFPRM-SBLFS. Among them, the ST-Acc indicator is calculated for nine flight modes and 10 other modes in total. The compared methods include classical neural networks that are widely used in various pattern recognition tasks.

[0146] The multi-strategy affine canonical timewarping approach (MACTW) is an extension of the affine canonical timewarping (ACTW), a new registration method for flight mode recognition. (Functional Data Analysis, FDA+C-Means) is a system based on the unsupervised learning paradigm. The system uses soft member classification technology and can even consider hybrid systems and transition systems. It is a new flight mode recognition method. SSGAN is a semi-supervised target trajectory algorithm based on the semi-supervised generative adversarial network (SSGAN). The algorithm can be used to identify abnormal targets.

[0147] We also provide LCLS loss and TIOU loss as loss functions for comparison. As shown in Table 2, IFPRM-SBLFS outperforms other competing methods in all indicators and achieves an overall recognition accuracy of 99.07%.

[0148] Table 3 Time complexity and space complexity of IFPRM-SBLFS

[0149] Table 3Time Complexity and Space Complexity of IFPRM-SBLFS

[0150]

[0151] In addition to model performance, the recognition speed of the method is also crucial in practical applications. As shown in Table 3, we evaluated the time complexity and space complexity of IFPRM-SBLFS. During the training phase, the average processing time of the model for a single sample is 1.72ms. The results show that the processing speed of this method can meet the application requirements in tasks with high flight real-time requirements. It is worth noting that in practical applications, when hardware performance is limited, the processing speed can be improved by adjusting the sampling interval.

[0152] 4.2 System continuity analysis

[0153] The above four indicators are used to quantitatively analyze the accuracy and performance of the prediction results. In order to intuitively demonstrate the advantages of the model in continuous prediction, Figure 3 The visualization results of the flight mode are provided. In this figure, four typical flight parameters, namely altitude, roll angle, pitch angle and speed, are selected for different flight states to highlight the characteristics of the flight process. Figure 3 In , each row represents the curve changes of the four key flight parameters under a single system, and the horizontal axis is the time axis. The vertical axis of the first column is the altitude, the vertical axis of the second column is the roll angle, the vertical axis of the third column is the pitch angle, and the vertical axis of the fourth column is the speed. In order to reduce the noise in the original signal, the data was preprocessed: first, the abnormal points such as spikes and omissions caused by acquisition errors were directly eliminated; secondly, the moving average smoothing algorithm was used to reduce the interference of environmental noise. In addition, in order to make different flight parameters within a similar numerical range, all signals are normalized according to the change range of each parameter to ensure that the model can learn the characteristics of each parameter. The blue curve in the figure represents the four flight parameter data collected in real time. The prediction results of the IFPRM-SBLFS model are marked with red circles, and the light purple background is used to distinguish a single continuous flight state. It should be pointed out in particular that the duration of different flight states follows a long-tail distribution. For ease of display, Figure 3The results in are appropriately scaled along the time axis.

[0154] During takeoff, as the pitch angle gradually increases, the aircraft gradually lifts off the ground, and the pressure altitude also increases. For climbs and rolls, the pressure altitude gradually increases as the roll angle changes 360°. For half-rolls, the pitch angle gradually returns to normal after the change, and the pressure altitude decreases as the roll angle changes. For lazy eight characters, the roll angle, pitch angle, pressure altitude, and indicated airspeed will change periodically throughout the process. For rolls, a 360° change in the roll angle is correctly judged as a roll process. During climbs and turns, the heading angle changes steadily and the pressure altitude gradually increases. In horizontal flight, the aircraft maintains the same heading angle and pressure altitude. During descending turns, the heading angle changes steadily and the pressure altitude gradually decreases. For landing, as the pressure altitude gradually decreases, the aircraft is approaching the ground. The results show that the prediction results can capture the core area of ​​each state and maintain strong local continuity. Faced with various states of different durations, the model can adaptively and accurately determine its flight trajectory and accurately match the starting and ending points of each state.

[0155] The IFPRM-SBLFS model also has potential in attitude prediction of different types of aircraft, but it is necessary to collect specific training data according to the target aircraft to ensure model adaptability and prediction accuracy. The data collection and annotation process is crucial. Different flight modes need to be manually annotated to generate clear flight mode boundaries for model deep learning training. During the training process, the model's boundary recognition ability and convergence are improved through methods such as data preprocessing and optimization of loss functions, thereby achieving accurate flight attitude prediction.

[0156] 5 Conclusion

[0157] This paper proposes an intelligent flight mode recognition method based on sensitive boundaries and long flight sequences to address the limitations of existing flight mode recognition methods in complex situations. By designing a specific model architecture and loss function, the IFPRM-SBLFS model can effectively identify flight states, accurately extract mode boundaries, and show high efficiency and accuracy in practical applications.

[0158] Experimental results show that the accuracy of IFPRM-SBLFS in flight status recognition is 99.23%, which fully proves its wide application potential in aviation safety monitoring. This invention effectively solves the problem of flight status recognition from the perspective of time detection for the first time. The proposed method combines sensitive boundaries with long flight sequences, overcomes the challenges of traditional methods in single and compound maneuver recognition, and can accurately locate temporal boundaries.

[0159] In view of the particularity of flight data, the present invention designs a highly automated specific architecture for IFPRM-SBLFS, which reduces the dependence on expert knowledge and experience, making it easy to realize online automatic recognition on aircraft. The adaptive graph embedding module designed in this model can adapt to different flight modes; the multi-scale transformer detector module is used to identify complex patterns; at the same time, the classification weighted focal loss and the regression joint spatiotemporal intersection loss are combined to effectively deal with the classification and regression losses of flight modes.

[0160] Through training and annotation of real flight data sets, IFPRM-SBLFS has shown significant advantages in precision and accuracy, further verifying its effectiveness in flight status monitoring tasks. The theory and method proposed in this paper provide a new technical path and innovative method for aviation safety monitoring.

[0161] The above is an example of the best implementation of the present invention, and the parts not described in detail are common knowledge of ordinary technicians in the field. The protection scope of the present invention shall be based on the content of the claims, and any equivalent transformation based on the technical enlightenment of the present invention is also within the protection scope of the present invention.

Claims

1. A flight mode intelligent recognition method based on IFPRM-SBLFS deep learning, characterized in that: The following steps are involved: Step 1: collect flight data, and manually annotate the flight mode of each time stage in each flight data to obtain training samples, and multiple training samples constitute a training set; Establish an IFPRM-SBLFS deep learning-based flight mode intelligent recognition model including an adaptive graph embedding model, a self-attention transformer module, a denoising deep multi-scale autoencoder, a boundary regressor, and a pattern classifier; Step 2: Input the training samples into the adaptive graph embedding model, extract the short-term and long-term flight parameter change trends respectively through sliding windows of different lengths, use spectral graph convolution to extract spatial features of different time scales and perform multi-scale feature splicing to obtain the comprehensive spatial features at each moment; Step 3: Input the comprehensive spatial features into the self-attention transformer module, analyze the short-term and long-term temporal dependencies through the self-attention transformer module and the denoising deep multi-scale autoencoder, and generate multi-scale features; Step 4: Input the multi-scale features into the boundary regressor and the mode classifier respectively, the mode classifier outputs the flight mode to which each time stage in the flight data belongs, and the boundary regressor outputs the boundary of the flight mode to which each time stage in the flight data belongs; Step 5: Compare the output results of the boundary regressor and the pattern classifier with the manual annotation results, calculate the classification weighted focal loss of the boundary regressor and the joint spatiotemporal intersection loss of the pattern classifier, and adjust the parameters of the boundary regressor and the pattern classifier through back propagation; Step 6: Repeat steps 1-5 to train the boundary regressor and pattern classifier by minimizing the loss function to obtain the trained boundary regressor and pattern classifier, and finally obtain the trained flight mode intelligent recognition model based on IFPRM-SBLFS deep learning; Step 7: Input the flight data of the flight to be identified into the trained flight mode intelligent recognition model based on IFPRM-SBLFS deep learning, and the model outputs the flight mode belonging to each time stage in the flight.

2. The flight mode intelligent recognition method based on IFPRM-SBLFS deep learning according to claim 1 is characterized in that: In step 1, the flight modes include take-off, level flight, climb, climb and rotation, sledgehammer, 360° roll, barrel turn, climb and roll, half roll, somersault, lazy eight, landing and rotation, dive, and landing. If none of the 11 modes are met, it is classified as "other" mode.

3. The flight mode intelligent identification method based on IFPRM-SBLFS deep learning according to claim 1 is characterized in that: In step 1, each flight data includes the real-time altitude, position, speed, roll angle and pitch angle data of the aircraft corresponding to all moments in a flight sortie.

4. The flight mode intelligent recognition method based on IFPRM-SBLFS deep learning according to claim 1 is characterized in that: In the adaptive graph embedding model, a multi-scale sliding window mechanism is adopted to extract the short-term and long-term flight parameter change trends through sliding windows of different lengths, and then integrate the short-term and long-term flight parameter change trends into the graph construction process. In the original adjacency matrix construction process, the Gaussian kernel function is introduced into the Euclidean distance to construct the adjacency matrix to enhance the description of nonlinear relationships. At the same time, the adjacency matrix is ​​adaptively weighted in combination with the attention mechanism. After extracting spatial features of different scales by spectral graph convolution, the graph convolution features of all scales are spliced ​​to obtain the comprehensive spatial features at each moment.

5. The flight mode intelligent identification method based on IFPRM-SBLFS deep learning according to claim 1 is characterized in that: In the self-attention transformer module, a transformer is used to replace the traditional short window filter. The self-attention transformer module takes the comprehensive spatial features as input, and weights and averages all the features at each moment to generate new features at each moment. Specifically, the weight of each feature during weighting is the correlation between the current feature and other features.

6. The flight mode intelligent identification method based on IFPRM-SBLFS deep learning according to claim 1 is characterized in that: The denoising deep multi-scale autoencoder includes an encoder, a decoder and a hidden layer, which is used to encode and decode the feature data output by the self-attention transformer module to obtain the potential features of the data, and output the corresponding reconstructed data through the decoder to generate multi-scale features.

7. The flight mode intelligent identification method based on IFPRM-SBLFS deep learning according to claim 1 is characterized in that: The network structure of the pattern classifier is expanded from the traditional three-layer one-dimensional convolutional network to a four-layer structure, and a multi-scale feature fusion module is introduced in each layer to enhance the network's ability to extract contextual information from different scales. By adopting diverse convolution kernel sizes, the network can simultaneously capture local and global features, and the output of each layer is enhanced through a multi-head attention mechanism, using eight attention heads to perform fine-grained allocation of feature weights.

8. The flight mode intelligent identification method based on IFPRM-SBLFS deep learning according to claim 1 is characterized in that: The boundary regressor adopts a four-layer network structure. Each layer is processed by a multi-scale fusion module with convolution kernel sizes of 3, 5, and 7. A context-based range constraint module is introduced in each layer to ensure that the boundary dynamic information in the features is effectively captured and processed. The output of each layer is enhanced by a multi-head attention mechanism to fully express the features at different scales. In each layer of the network, in addition to the ReLU activation function, a Leaky ReLU is also introduced. The output range of the boundary regressor is adaptively adjusted according to the training data, where the first layer maintains the initial range and the subsequent layers dynamically expand the range according to the data distribution.

Images