Two-stage flight parameter data abnormity identification method, device, equipment and medium

Abstract

The invention relates to the technical field of avionics, discloses a dual-stage flight parameter data anomaly identification method, device and equipment and a medium, breaks through the limitation of traditional single-dimensional threshold detection, can adapt to dynamic characteristic changes of different flight stages, and improves the detection accuracy. Particularly, the method has higher recognition sensitivity for tiny anomalies in a complex maneuvering action period, and meanwhile, the objectivity of model evaluation is ensured through strict division of a training set and a test set. The problems that current flight parameter data fault diagnosis is highly dependent on manpower, poor in static criterion adaptability and poor in interpretation precision are solved through a double-stage recognition strategy, high efficiency and intelligence of flight parameter data fault diagnosis are achieved, and accuracy and robustness of flight parameter data anomaly detection are improved.

Description

Technical Field

[0001] This invention relates to the field of avionics technology and discloses a method, apparatus, equipment, and medium for identifying anomalies in two-stage flight parameter data. Background Technology

[0002] With the continuous development of modern aviation equipment, flight safety analysis and condition monitoring have become increasingly prominent issues in the face of diverse missions and complex operating environments. Flight parameter data is a type of multidimensional time-series data, containing all the key information about each flight mission performed by the aircraft. Therefore, the key to aircraft condition monitoring and health management lies in the detection of flight parameter data anomalies, and quickly identifying flight parameter data belonging to system-related faults among all abnormal flight parameter data is of paramount importance.

[0003] Current methods for detecting anomalies in flight parameter data primarily involve defining a specific threshold for each indicator. When the corresponding indicator in the flight parameter data exceeds the threshold range, the system automatically identifies an anomaly. This method requires extensive experience from senior experts and cannot perform multi-parameter joint analysis. Aircraft faults generally exhibit characteristics of being covert, gradual, and having strong correlations among multiple parameters. Therefore, when detecting faults in flight parameter data, the multidimensional correlations of different parameters should be considered. Furthermore, the threshold for judging flight parameter data anomalies varies depending on the specific flight maneuvers the aircraft performs. Therefore, flight parameter data screening based on flight maneuvers is crucial for troubleshooting flight parameter data. Summary of the Invention

[0004] The purpose of this invention is to provide a two-stage flight parameter data anomaly identification method, device, equipment, and medium to solve the problems of current flight parameter data fault diagnosis, such as strong reliance on manual intervention, poor adaptability of static criteria, and poor interpretation accuracy, thereby achieving efficient and intelligent flight parameter data fault diagnosis.

[0005] To achieve the above-mentioned technical effects, the technical solution adopted by the present invention is as follows: A two-stage flight parameter data anomaly identification method includes: S1. Collect raw flight parameter data of historical aircraft sorties and sorties to be identified, and extract flight parameter subsequences related to flight maneuvers from the raw flight parameter data. The raw flight parameter data includes multi-dimensional time-series data of relative pressure altitude, climb rate, radio altimeter indicated altitude, heading angle, pitch angle, airspeed, and normal acceleration. S2. The flight parameter subsequence is converted into a short-time Fourier transform (SFT) time-frequency map using SFT, and each SFT time-frequency map is labeled with a flight action category to form the aircraft's flight action dataset. The SFT time-frequency maps corresponding to historical sorties in the flight action dataset are divided into a training set, and the SFT time-frequency maps corresponding to the sorties to be identified are divided into a test set. The flight action categories include 180° turn, climb, descent, incorrect control, and maximum angular velocity turn. S3. Construct a deep learning neural network model for recognizing flight action categories. Use the training set in the flight action dataset to train the model, ensuring that the loss function value is within a preset range, and obtain the trained deep learning neural network model. S4. Using the test set in the flight action dataset as input, the trained deep learning neural network model is used to classify the flight action category of the short-time Fourier time-frequency image in the test set to obtain the classified short-time Fourier time-frequency image. S5. Establish a single-dimensional original flight parameter data anomaly detection model based on dynamic time-warping algorithm, and use the anomaly detection model to generate reference samples corresponding to each flight action category; calculate the DTW distance between the flight parameter subsequence to be identified and the corresponding reference sample corresponding to the short-time Fourier time-frequency graph after each classification; if the DTW distance is outside the dynamic preset threshold range, it is determined that the flight parameter data of the aircraft to be identified is abnormal.

[0006] Further, step S1 includes the following steps: Based on the original flight parameter data, determine the normal load coefficient at each time point in the original flight parameter data; Starting from the time starting point of the original flight parameter data, when the normal load coefficient at a certain time point is greater than or equal to the preset normal overload threshold, it is determined that the aircraft is performing a certain flight maneuver. Then, the time points where the normal load coefficient is 1 are searched forward and backward, and combined with the points where the roll angle and pitch angular velocity approach zero, the entry point and recovery point of the aircraft's flight maneuver are determined. The original flight parameter data within the time period from the aircraft entering the flight maneuver to the recovery point of the flight maneuver are segmented and used as the flight parameter subsequence of the aircraft performing the flight maneuver. The original flight parameter data is traversed until the segmentation of the entire original flight parameter data is completed.

[0007] Further, step S2 includes the following steps: The flight parameter subsequence is divided into several frame time segments according to a preset window length; Each time segment is multiplied by a Hamming window function, and then the single-sided amplitude spectrum of each frame is extracted using short-time Fourier transform. The single-sided amplitude spectra of each frame are arranged in chronological order to form a two-dimensional time-frequency matrix. Based on the two-dimensional time-frequency matrix, a short-time Fourier time-frequency diagram is generated through coordinate mapping.

[0008] Furthermore, the expression for the short-time Fourier transform is:

[0009] in, This is the short-time Fourier output result. For flight parameter subsequence, For the first flight parameter subsequence The characteristic flight parameter signal, For window functions, For the center time of the window, It is a complex exponential function.

[0010] Furthermore, the window function is a Hamming window function, expressed as: ; in: For window functions, For window length, For time indexing, .

[0011] Further, step S5 includes the following steps: Based on the normal flight parameter subsequences corresponding to the short-time Fourier time-frequency plots of different historical sorties in the test set, one of the flight actions is selected as the target action. The DTW distance between the normal flight parameter subsequences belonging to the target action in each sortie and the normal flight parameter subsequences belonging to the target action in other sorties is calculated. Then, the total DTW distance between the normal flight parameter subsequences belonging to the target action in each sortie and the normal flight parameter subsequences belonging to the target action in other sorties is calculated. Select the sortie with the smallest total DTW distance, and use the normal flight parameter subsequence of that sortie as a reference sample for the target flight action; For all historical sorties belonging to the target maneuver's normal flight parameter subsequences, calculate the mean of all DTW distances. and standard deviation The mean DTW distance for each flight maneuver is determined by traversing all historical sorties and the skewness and kurtosis of all DTW distance distributions. and standard deviation , and the skewness and kurtosis of all DTW distance distributions; Based on the skewness and kurtosis of all DTW distance distributions for each flight maneuver, through Calculate the dynamic coefficients for each flight maneuver, where, For the dynamic coefficient of flight maneuvers, , For preset adjustment coefficients, The skewness of the DTW distance distribution for all flight maneuvers. Kuness of the DTW distance distribution for all flight maneuvers; For the flight parameter subsequence to be identified corresponding to the short-time Fourier time-frequency graph of each flight action category in the test set, calculate the DTW distance between the flight parameter subsequence to be identified and the corresponding reference sample. If the DTW distance is within... If the data is outside the specified range, the flight parameter data of the aircraft to be identified is determined to be abnormal.

[0012] A two-stage flight parameter data anomaly identification device, used to implement the aforementioned two-stage flight parameter data anomaly identification method, includes: The flight parameter data acquisition module is used to collect raw flight parameter data of historical sorties and sorties to be identified, and to extract flight parameter subsequences related to flight maneuvers from the raw flight parameter data. The raw flight parameter data is multi-dimensional time-series data including relative pressure altitude, climb rate, radio altimeter indicated altitude, heading angle, pitch angle, airspeed and normal acceleration. The flight parameter data processing module is used to convert the flight parameter subsequence into a short-time Fourier transform (SFT) time-frequency map using SFT, and to label each SFT time-frequency map with flight maneuver categories to form a flight maneuver dataset for the aircraft. The SFT time-frequency maps corresponding to historical sorties in the flight maneuver dataset are divided into a training set, and the SFT time-frequency maps corresponding to the sorties to be identified are divided into a test set. The flight maneuver categories include 180° turn, climb, descent, incorrect control, and maximum angular velocity turn. The deep learning neural network model building module is used to build a deep learning neural network model for recognizing flight action categories. It uses the training set in the flight action dataset to train the model, ensuring that the loss function value is within a preset range, and obtains the trained deep learning neural network model. The flight action classification module is used to take the test set in the flight action dataset as input, and use the trained deep learning neural network model to classify the flight action category of the short-time Fourier time-frequency image in the test set to obtain the classified short-time Fourier time-frequency image. The abnormal data detection module is used to establish an anomaly detection model for single-dimensional raw flight parameter data based on the dynamic time-warping algorithm, and to generate reference samples corresponding to each flight action category using the anomaly detection model; calculate the DTW distance between the flight parameter subsequence to be identified and the corresponding reference sample corresponding to the short-time Fourier time-frequency graph after each classification; if the DTW distance is outside the dynamic preset threshold range, it is determined that the flight parameter data of the aircraft to be identified is abnormal.

[0013] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the aforementioned two-stage flight parameter data anomaly identification method.

[0014] A computer-readable storage medium storing a computer program that performs the aforementioned two-stage flight parameter data anomaly identification method.

[0015] Compared with the prior art, the beneficial effects of this invention are: This invention first utilizes a deep learning neural network model to accurately classify flight maneuvers in the first stage, transforming time-series data into time-frequency feature maps through short-time Fourier transform, effectively capturing the dynamic characteristics of flight maneuvers. Secondly, in the second stage, a dynamic time-normalization algorithm model is constructed for different flight maneuver categories, generating a reference sample library highly matched to actual flight states. Anomaly detection is achieved by calculating the DTW distance between test samples and reference samples. This method overcomes the limitations of traditional single-dimensional threshold detection, adapting to changes in dynamic characteristics at different flight stages, and exhibiting higher sensitivity in recognizing minute anomalies, especially during complex maneuvers. Furthermore, the strict division of training and test sets ensures the objectivity of model evaluation. This two-stage identification strategy addresses the current problems of heavy reliance on manual intervention, poor adaptability to static criteria, and low interpretation accuracy in flight parameter data fault diagnosis, achieving efficient and intelligent fault diagnosis of flight parameter data and improving the accuracy and robustness of flight parameter data anomaly detection. Attached Figure Description

[0016] Figure 1 This is a flowchart of the two-stage flight parameter data anomaly identification method in the embodiment; Figure 2 This is a schematic diagram of the architecture of the deep learning neural network model in the embodiment; Figure 3 This is a block diagram of the architecture of the two-stage flight parameter data anomaly identification device in the embodiment; Figure 4 This is a structural block diagram of the computer device in the embodiment; Wherein, 1-memory, 2-processor. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings. However, this should not be construed as limiting the scope of the above-described subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0018] Example 1 This embodiment provides a method for identifying anomalies in two-stage flight parameter data, including: S1. Collect raw flight parameter data of historical aircraft sorties and sorties to be identified, and extract flight parameter subsequences related to flight maneuvers from the raw flight parameter data. The raw flight parameter data includes multi-dimensional time-series data of relative pressure altitude, climb rate, radio altimeter indicated altitude, heading angle, pitch angle, airspeed, and normal acceleration. S2. The flight parameter subsequence is converted into a short-time Fourier transform (SFT) time-frequency map using SFT, and each SFT time-frequency map is labeled with a flight action category to form the aircraft's flight action dataset. The SFT time-frequency maps corresponding to historical sorties in the flight action dataset are divided into a training set, and the SFT time-frequency maps corresponding to the sorties to be identified are divided into a test set. The flight action categories include 180° turn, climb, descent, incorrect control, and maximum angular velocity turn. S3. Construct a deep learning neural network model for recognizing flight action categories. Use the training set in the flight action dataset to train the model, ensuring that the loss function value is within a preset range, and obtain the trained deep learning neural network model. S4. Using the test set in the flight action dataset as input, the trained deep learning neural network model is used to classify the flight action category of the short-time Fourier time-frequency image in the test set to obtain the classified short-time Fourier time-frequency image. S5. Establish a single-dimensional original flight parameter data anomaly detection model based on dynamic time-warping algorithm, and use the anomaly detection model to generate reference samples corresponding to each flight action category; calculate the DTW distance between the flight parameter subsequence to be identified and the corresponding reference sample corresponding to the short-time Fourier time-frequency graph after each classification; if the DTW distance is outside the dynamic preset threshold range, it is determined that the flight parameter data of the aircraft to be identified is abnormal.

[0019] This invention first utilizes a deep learning neural network model to accurately classify flight maneuvers in the first stage, transforming time-series data into time-frequency feature maps through short-time Fourier transform, effectively capturing the dynamic characteristics of flight maneuvers. Secondly, in the second stage, a dynamic time-normalization algorithm model is constructed for different flight maneuver categories, generating a reference sample library highly matched to actual flight states. Anomaly detection is achieved by calculating the DTW distance between test samples and reference samples. This method overcomes the limitations of traditional single-dimensional threshold detection, adapting to changes in dynamic characteristics at different flight stages, and exhibiting higher sensitivity in recognizing minute anomalies, especially during complex maneuvers. Furthermore, the strict division of training and test sets ensures the objectivity of model evaluation. This two-stage identification strategy addresses the current problems of heavy reliance on manual intervention, poor adaptability to static criteria, and low interpretation accuracy in flight parameter data fault diagnosis, achieving efficient and intelligent fault diagnosis of flight parameter data and improving the accuracy and robustness of flight parameter data anomaly detection.

[0020] Based on the same inventive concept, see [link to inventive concept] Figure 3 This embodiment also provides a two-stage flight parameter data anomaly identification device for the aforementioned two-stage flight parameter data anomaly identification method, comprising: The flight parameter data acquisition module is used to collect raw flight parameter data of historical sorties and sorties to be identified, and to extract flight parameter subsequences related to flight maneuvers from the raw flight parameter data. The raw flight parameter data is multi-dimensional time-series data including relative pressure altitude, climb rate, radio altimeter indicated altitude, heading angle, pitch angle, airspeed and normal acceleration. The flight parameter data processing module is used to convert the flight parameter subsequence into a short-time Fourier transform (SFT) time-frequency map using SFT, and to label each SFT time-frequency map with flight maneuver categories to form a flight maneuver dataset for the aircraft. The SFT time-frequency maps corresponding to historical sorties in the flight maneuver dataset are divided into a training set, and the SFT time-frequency maps corresponding to the sorties to be identified are divided into a test set. The flight maneuver categories include 180° turn, climb, descent, incorrect control, and maximum angular velocity turn. The deep learning neural network model building module is used to build a deep learning neural network model for recognizing flight action categories. It uses the training set in the flight action dataset to train the model, ensuring that the loss function value is within a preset range, and obtains the trained deep learning neural network model. The flight action classification module is used to take the test set in the flight action dataset as input, and use the trained deep learning neural network model to classify the flight action category of the short-time Fourier time-frequency image in the test set to obtain the classified short-time Fourier time-frequency image. The abnormal data detection module is used to establish an anomaly detection model for single-dimensional raw flight parameter data based on the dynamic time-warping algorithm, and to generate reference samples corresponding to each flight action category using the anomaly detection model; calculate the DTW distance between the flight parameter subsequence to be identified and the corresponding reference sample corresponding to the short-time Fourier time-frequency graph after each classification; if the DTW distance is outside the dynamic preset threshold range, it is determined that the flight parameter data of the aircraft to be identified is abnormal.

[0021] Example 2 See Figure 1 This embodiment further details the method for identifying anomalies in two-stage flight parameter data, and the specific steps include: S1. Collect raw flight parameter data of historical aircraft sorties and sorties to be identified, and extract flight parameter subsequences related to flight maneuvers from the raw flight parameter data. The raw flight parameter data includes multi-dimensional time-series data of relative pressure altitude, climb rate, radio altimeter indicated altitude, heading angle, pitch angle, airspeed, and normal acceleration.

[0022] Specifically, firstly, basic information and raw flight parameter data of the aircraft's historical sorties and sorties to be identified are collected. The basic information includes, but is not limited to, key attributes such as aircraft type and sortie batch; the raw flight parameter data is formatted and encoded multi-dimensional time-series data collected continuously in chronological order, which is a long sequence and specifically includes key parameters under various flight conditions, such as relative pressure altitude, climb rate, radio altimeter altitude, heading angle, pitch angle, airspeed, and normal acceleration.

[0023] Then, the original flight parameter data is coarsely segmented according to a preset normal overload threshold to extract the flight parameter subsequences for aircraft maneuvers. Based on the calculation formulas for normal acceleration and normal load coefficient, the normal load coefficient at each time point in the original flight parameter data is determined. According to the flight manual, the normal overload threshold is set to N=1.5, which serves as the segmentation index for coarse segmentation of aircraft maneuvers. Starting from the time starting point of the original flight parameter data, coarse segmentation is performed. During segmentation, if the normal load coefficient of the aircraft at time t is greater than or equal to the normal overload threshold of 1.5 (the normal load coefficient is 1 during level flight), it is determined that the aircraft is performing a certain flight maneuver at time t. Using time t as the base point, by searching forward and backward for time points where the normal load coefficient is 1 and combining this with time points where roll angle and pitch angular velocity approach zero, the time point when the aircraft enters the flight maneuver and the time point when the aircraft recovers from the flight maneuver are determined. The original flight parameter data within the corresponding time period from the aircraft entering the flight maneuver to recovering from the flight maneuver are taken as the flight parameter subsequence for the aircraft performing that flight maneuver. Finally, repeat the above segmentation operation until the entire sequence of original flight parameter data is segmented. For the several flight parameter subsequences obtained in advance, if the interval between adjacent flight parameter subsequences is less than 1 second, they are merged into one flight parameter subsequence, representing the same flight action; if the flight parameter subsequence has problems such as the flight action duration exceeding the preset limit range and parameter contradictions, the flight parameter subsequence is identified as an abnormal sequence and removed, and finally several flight parameter subsequences corresponding to different flight actions are extracted.

[0024] S2. The flight parameter subsequence is converted into a short-time Fourier transform (SFT) time-frequency graph using the short-time Fourier transform (SFT), and each SFT time-frequency graph is labeled with a flight action category to form the flight action dataset of the aircraft. The flight action categories include 180° turn, ascent, descent, mishandling, and maximum angular velocity turn.

[0025] Short-time Fourier transform (SFT) is a multi-scale time-frequency analysis method. It obtains the joint distribution of flight parameter data in time and frequency by locally truncating flight parameter data using a sliding window function on the time axis and performing a Fourier transform on each truncated segment. Compared to one-dimensional signals, time-frequency graphs carry richer information and can further enhance the flight parameter characteristic information of time series. Therefore, this invention uses SFT for the coarsely segmented subsequences of different flight maneuvers. The SFT expression is:

[0026] in, This is the short-time Fourier output result; For flight parameter subsequence; For the first flight parameter subsequence Characteristic flight parameter signals, such as relative pressure altitude, rate of climb and fall, radio altimeter reading altitude, heading angle, pitch angle, airspeed, and normal acceleration; It is a window function used to extract local segments of a signal, and has time-frequency localization characteristics; The center time of the window represents the current analysis time point; It is a complex exponential function that implements the projection from the time-frequency domain to the frequency domain.

[0027] Based on the characteristics of flight parameter data, this invention selects the Hamming window, which has high frequency resolution and is suitable for capturing transient signals or rapidly changing data, as the window function. The expression is as follows: ; in: For window functions, For window length, For time indexing, The window length of this invention It is 256.

[0028] Specifically, the process of converting the flight parameter subsequences into short-time Fourier transform time-frequency maps using short-time Fourier transform, and labeling each short-time Fourier transform time-frequency map with flight action categories to form the aircraft's flight action dataset includes the following steps: First, for each flight parameter subsequence, the flight parameter data is divided into multiple consecutive time segments with a preset window length of 256. Each time segment is a frame. To ensure the continuity of the time dimension, adjacent time segments slide by using the method of "step size = window length - number of overlapping points" to avoid data gaps. Then, each frame time segment is multiplied by a Hamming window function to suppress spectral leakage.

[0029] Secondly, a short-time Fourier transform is performed on each frame of time segment data after multiplication by a Hamming window to convert the time-domain signal of each flight parameter subsequence into a frequency-domain signal, and the single-sided amplitude spectrum of each frame is extracted. The extracted single-sided amplitude spectra of each frame are then arranged in chronological order to form a two-dimensional time-frequency matrix. The behavior frequency of the two-dimensional time-frequency matrix is ​​listed as time, and the visualization contrast is enhanced.

[0030] Finally, the two-dimensional time-frequency matrix is ​​transformed into a short-time Fourier time-frequency graph through coordinate mapping and visualized. Then, based on the actual flight maneuvers corresponding to each flight parameter subsequence, such as 180° turns, climbs, and maximum angular velocity turns, the short-time Fourier time-frequency graphs corresponding to each flight parameter subsequence are labeled with flight maneuver categories to obtain the aircraft's flight maneuver dataset. The flight maneuver categories include 180° turns, climbs, descents, mishandling, and maximum angular velocity turns.

[0031] It should be noted that the short-time Fourier time-frequency graphs corresponding to historical sorties in the flight action dataset are divided into the training set, and the short-time Fourier time-frequency graphs corresponding to the sorties to be identified are divided into the test set.

[0032] S3. Construct a deep learning neural network model for recognizing flight action categories. Use the training set in the flight action dataset to train the model, ensuring that the loss function value is within a preset range, and obtain the trained deep learning neural network model.

[0033] Specifically, the constructed deep learning neural network model includes an input layer, a feature extraction layer, and a softmax output layer, see [link to documentation]. Figure 2 .

[0034] The input layer comprises two consecutive convolutional layers with 16 and 32 convolutional kernels respectively, ReLU activation, and max pooling. The input layer receives the 224×224×3 multi-dimensional short-time Fourier time-frequency map output from step S2. After convolution, BatchNorm2d batch normalization, ReLU activation, and max pooling, it outputs a preliminary feature map. , , This is the number of channels in the initial feature map. , The initial feature map consists of its height and width.

[0035] The feature extraction module consists of a dynamically separable convolution module and a CBAM attention mechanism module. The dynamically separable convolution is used for local feature extraction from the short-time Fourier time-frequency image, while the CBAM attention mechanism enhances the expressive power of the convolutional network through adaptive feature refinement.

[0036] The feature extraction module uses a preliminary feature map As input, the dynamically separable convolutional module first performs a 1×1 convolution on the initial feature map. Adjust the channel dimensions: Obtain intermediate feature maps , , It is an intermediate feature map The number of channels, , The middle feature map is divided into Height and width. (Through) Perform cross-channel convolution to increase the number of channels from Mapped to This provides a unified input dimension for subsequent multi-scale convolutions; then, the intermediate feature maps are processed... Each channel is applied with four different scale depth kernels ( The system generates four distinct branches. The outputs of these four branches are then concatenated along the channel dimension, and a 1×1 convolution is used to compress the channels, restoring the original dimensions. The time-frequency feature map is obtained by enhancing nonlinearity through the GELU activation function and batch normalization. , H and W are time-frequency feature maps, respectively. Height and width, Time-frequency feature map The number of channels.

[0037] The CBAM attention mechanism module consists of a channel attention module and a spatial attention module, enabling multi-dimensional feature enhancement of the flight parameter time-frequency feature map. Specifically, the channel attention module uses the time-frequency feature map... As input, first analyze the time-frequency feature map of each channel. Take the average value and maximum value Two pooling results were obtained respectively. , Combine the two pooling results , The shared multilayer perceptron (MLP) is used to learn the non-linear dependencies between channels and generate channel attention weights. , Finally, the channel attention weights are... Time-frequency feature map Multiply channel by channel to output the time-frequency feature map after channel attention weighting. The expression is , .

[0038] The spatial attention module uses the feature map weighted by channel attention. As input, first process the feature map of each channel. Take the average value and maximum value Two pooling results were obtained respectively. , Then the two pooling results , Stitch along the channel dimension, then pass through Convolution compresses the number of channels to 1.

[0039] Finally, spatial attention weights are... Time-frequency feature map Multiplying by spatial location yields the final output high-dimensional feature map. The expression is .

[0040] The output layer consists of a fully connected layer and a softmax function. The fully connected layer extracts high-dimensional feature maps. Mapping to the flight action category dimension to construct a high-dimensional feature map The association between the fully connected layer and the aircraft action category is then established. The Softmax function converts the output of the fully connected layer into the posterior probability distribution of each aircraft action category, which facilitates the accurate identification of different aircraft actions.

[0041] The deep learning neural network model is trained and its parameters are tuned using the training set obtained in step S2. In this embodiment, the training set loss is calculated using a loss function, and the parameters of the model are optimized and updated using a label smoothing regularization function and the Adam optimizer, so that the deep learning neural network model reaches its optimal state.

[0042] S4. Using the test set in the flight action dataset as input, the trained deep learning neural network model is used to classify the short-time Fourier time-frequency maps in the test set according to the first stage of flight action category classification, and the classified short-time Fourier time-frequency maps are obtained. That is, the short-time Fourier time-frequency maps in the test set are classified according to five flight action categories: 180° turn, ascent, descent, incorrect control, and maximum angular velocity turn.

[0043] S5. Establish a single-dimensional original flight parameter data anomaly detection model based on dynamic time-warping algorithm, and use the anomaly detection model to generate reference samples corresponding to each flight action category; calculate the DTW distance between the flight parameter subsequence to be identified and the corresponding reference sample corresponding to the short-time Fourier time-frequency graph after each classification; if the DTW distance is outside the dynamic preset threshold range, it is determined that the flight parameter data of the aircraft to be identified is abnormal.

[0044] Aircraft exhibiting similar flight parameter data curves under identical maneuvers generally exhibit similar patterns. This invention establishes a second-stage flight parameter data anomaly detection model based on the Dynamic Time Induction (DTW) algorithm to determine the dynamic threshold for flight parameter data anomalies. This method can dynamically adjust the limits as evaluation data accumulates. Specific steps include: S51: Based on the normal flight parameter subsequences corresponding to the short-time Fourier time-frequency diagrams of different historical sorties in the test set, select one of the flight actions as the target action, calculate the DTW distance between the normal flight parameter subsequences belonging to the target action in each sortie and the normal flight parameter subsequences belonging to the target action in other sorties, and then calculate the total DTW distance from the normal flight parameter subsequences belonging to the target action in each sortie to the normal flight parameter subsequences belonging to the target action in other sorties.

[0045] For example, assuming there are three normal flight parameter subsequences corresponding to historical sorties, and a 180° turn is selected as the target maneuver, firstly, calculate the DTW distance I between the normal flight parameter subsequences of sortie 1 and sortie 2 that belong to the 180° turn; calculate the DTW distance II between the normal flight parameter subsequences of sortie 1 and sortie 3 that belong to the 180° turn; and calculate the DTW distance II between the normal flight parameter subsequences of sortie 2 and sortie 3 that belong to the 180° turn. The DTW distance III between normal flight parameter subsequences of a 180° turn is calculated; then, the total DTW distance from the normal flight parameter subsequence of a 180° turn in sortie 1 to the normal flight parameter subsequence of the target action in other sorties is calculated as DTW distance I + DTW distance II. Similarly, the total DTW distance corresponding to the normal flight parameter subsequence of a 180° turn in sortie 2 is DTW distance I + DTW distance III, and the total DTW distance corresponding to the normal flight parameter subsequence of a 180° turn in sortie 3 is DTW distance II + DTW distance III.

[0046] S52: Select the sortie with the smallest total DTW distance, and use the normal flight parameter subsequence of the sortie as the reference sample of the target flight action. For example, for the target action of a 180° turn, assuming that the value of DTW distance I + DTW distance II is the smallest, then use the normal flight parameter subsequence of sortie 1 that belongs to the 180° turn as the reference sample of the 180° turn action. Traverse all normal flight parameter subsequences corresponding to flight maneuvers to determine the reference sample for each flight maneuver.

[0047] It should be noted that when calculating the total DTW distance corresponding to a certain flight maneuver for each sortie, if the lengths of the normal flight parameter subsequences corresponding to different sorties are inconsistent, the following method is used: The total DTW distance is normalized to eliminate the influence of length differences and avoid increased cumulative error in long sequences due to a large number of path steps. For a single sortie, the total DTW distance corresponds to a specific flight maneuver. The length of the normal flight parameter subsequence. This represents the normalized total DTW distance. When the normal flight parameter subsequence lengths corresponding to different sorties are inconsistent, the normalized total DTW distance is compared to determine the reference sample.

[0048] S53: For all normal flight parameter subsequences belonging to the target maneuver in all historical sorties, calculate the mean of all DTW distances. and standard deviation The mean DTW distance for each flight maneuver is determined by traversing all historical sorties and the skewness and kurtosis of all DTW distance distributions. and standard deviation And the skewness and kurtosis of all DTW distance distributions; for example, for a target maneuver of a 180° turn, calculate the mean of all DTW distances for three historical flights. and standard deviation and all DTW distance distributions and kurtosis The same applies to other flight maneuvers.

[0049] S54: Skewness of all DTW distance distributions for each flight maneuver and kurtosis ,pass Calculate the dynamic coefficients for each flight maneuver. In the formula, A and B are preset adjustment coefficients, which are manually adjusted and set according to actual needs. For example, for a 180° turn flight maneuver, the dynamic coefficient corresponding to the 180° turn flight maneuver is calculated. The same applies to other flight maneuvers.

[0050] S55: For the flight parameter subsequence to be identified corresponding to the short-time Fourier time-frequency graph of each flight action category in the test set, calculate the DTW distance between the flight parameter subsequence to be identified and the corresponding reference sample. If the DTW distance is within... If the data is outside the specified range, the flight parameter data of the aircraft to be identified is determined to be abnormal. For example, for the short-time Fourier time-frequency diagram classified as a 180° turn in the test set, a 180° turn flight maneuver reference sample can be obtained according to step S53, and the mean of all DTW distances corresponding to all historical flight parameter subsequences of the 180° turn flight maneuver can be determined according to step S54. and standard deviation Determine the corresponding dynamic coefficients according to step S55. Then, the DTW distance between the flight parameter subsequence to be identified corresponding to the short-time Fourier time-frequency diagram and the 180° turn flight action reference sample is calculated. If the DTW distance is within... If the data is outside the specified range, the flight parameter data for that particular flight is considered abnormal.

[0051] S56: To improve the robustness of flight parameter data anomaly detection, an adaptive threshold method is used to dynamically update the mean. Specific implementation: A sliding window method is used to calculate the mean and standard deviation of the DTW distance for the N most recent flights of the same aircraft with the same maneuver. The window is updated as new data is added.

[0052] This invention provides a coarse segmentation method for flight maneuvers, which converts multidimensional time-series flight parameter data into multidimensional short-time Fourier time-frequency graphs, effectively solving the technical problem of inconsistent flight parameter sequence lengths and providing standardized input for subsequent deep learning model construction.

[0053] This invention designs a network model for motion recognition, which mainly consists of a backbone extraction network, a multi-scale feature fusion module, and an attention mechanism. Compared with existing excellent networks, this network has better motion recognition accuracy.

[0054] This invention proposes a method for identifying flight parameter data anomalies based on DTW (Dynamic Flight Wire). This method categorizes the flight parameter subsequences after aircraft action identification according to different actions, and compares and analyzes the flight parameter subsequences under each category with dynamically generated standard flight parameter data, thereby realizing automated anomaly detection of the measured flight parameter data.

[0055] The method proposed in this invention consists of two stages. The first stage involves converting multidimensional flight parameter data into a multidimensional time series graph using a short-time Fourier transform (SFT). Then, a deep learning model is used to identify and group the flight parameter time series graph based on aircraft movements. The second stage employs the Time-Depth Winding (DTW) algorithm to detect anomalies in the grouped flight parameter data for different movements. This invention isolates differences in aircraft movements through movement classification, improves the comparability of flight parameter data for similar movements, continuously updates the DTW threshold, adapts to aircraft aging and retrofitting characteristics, and achieves rapid and accurate automated interpretation of flight parameter data.

[0056] The present invention also provides a computer device, such as Figure 4 As shown, the device includes a memory 1, a processor 2, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned two-stage flight parameter data anomaly identification method. Specifically, the computer device can be a computer terminal, a server, or a similar computing device.

[0057] In this embodiment, a computer-readable storage medium is provided, which stores a computer program that executes the above-described two-stage flight parameter data anomaly identification method.

[0058] Specifically, computer-readable storage media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer-readable storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable storage media do not include transient computer-readable media, such as modulated data signals and carrier waves.

[0059] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying anomalies in two-stage flight parameter data, characterized in that, include: S1. Collect raw flight parameter data of historical aircraft sorties and sorties to be identified, and extract flight parameter subsequences related to flight maneuvers from the raw flight parameter data. The raw flight parameter data includes multi-dimensional time-series data of relative pressure altitude, climb rate, radio altimeter indicated altitude, heading angle, pitch angle, airspeed, and normal acceleration. S2. The flight parameter subsequence is converted into a short-time Fourier transform (SFT) time-frequency map using SFT, and each SFT time-frequency map is labeled with a flight action category to form the aircraft's flight action dataset. The SFT time-frequency maps corresponding to historical sorties in the flight action dataset are divided into a training set, and the SFT time-frequency maps corresponding to the sorties to be identified are divided into a test set. The flight action categories include 180° turn, climb, descent, incorrect control, and maximum angular velocity turn. S3. Construct a deep learning neural network model for recognizing flight action categories. Use the training set in the flight action dataset to train the model, ensuring that the loss function value is within a preset range, and obtain the trained deep learning neural network model. S4. Using the test set in the flight action dataset as input, the trained deep learning neural network model is used to classify the flight action category of the short-time Fourier time-frequency image in the test set to obtain the classified short-time Fourier time-frequency image. S5. Establish a single-dimensional original flight parameter data anomaly detection model based on dynamic time-warping algorithm, and use the anomaly detection model to generate reference samples corresponding to each flight action category; calculate the DTW distance between the flight parameter subsequence to be identified and the corresponding reference sample corresponding to the short-time Fourier time-frequency graph after each classification; if the DTW distance is outside the dynamic preset threshold range, it is determined that the flight parameter data of the aircraft to be identified is abnormal.

2. The dual-stage flight parameter data anomaly identification method according to claim 1, characterized in that, Step S1 includes the following steps: Based on the original flight parameter data, determine the normal load coefficient at each time point in the original flight parameter data; Starting from the time starting point of the original flight parameter data, when the normal load coefficient at a certain time point is greater than or equal to the preset normal overload threshold, it is determined that the aircraft is performing a certain flight maneuver. Then, the time points where the normal load coefficient is 1 are searched from front to back, and combined with the time points where the roll angle and pitch angular velocity approach zero, the entry point and recovery point of the aircraft's flight maneuver are determined. The original flight parameter data within the time period from the aircraft entering the flight maneuver to the recovery point of the flight maneuver are segmented and used as the flight parameter subsequence of the aircraft performing the flight maneuver. The original flight parameter data is traversed until the segmentation of the entire original flight parameter data is completed.

3. The dual-stage flight parameter data anomaly identification method according to claim 2, characterized in that, Step S2 includes the following steps: The flight parameter subsequence is divided into several frame time segments according to a preset window length; Each time segment is multiplied by a Hamming window function, and then the single-sided amplitude spectrum of each frame is extracted using short-time Fourier transform. The single-sided amplitude spectra of each frame are arranged in chronological order to form a two-dimensional time-frequency matrix. Based on the two-dimensional time-frequency matrix, a short-time Fourier time-frequency diagram is generated through coordinate mapping.

4. The dual-stage flight parameter data anomaly identification method according to claim 3, characterized in that, The expression for the short-time Fourier transform is: in, This is the short-time Fourier output result. For flight parameter subsequence, For the first flight parameter subsequence The characteristic flight parameter signal, For window functions, For the center time of the window, It is a complex exponential function.

5. The dual-stage flight parameter data anomaly identification method according to claim 4, characterized in that, The window function used is the Hamming window function, and its expression is: ; in: For window functions, For window length, For time indexing, .

6. The dual-stage flight parameter data anomaly identification method according to claim 1, characterized in that, Step S5 includes the following steps: Based on the normal flight parameter subsequences corresponding to the short-time Fourier time-frequency plots of different historical sorties in the test set, one of the flight actions is selected as the target action. The DTW distance between the normal flight parameter subsequences belonging to the target action in each sortie and the normal flight parameter subsequences belonging to the target action in other sorties is calculated. Then, the total DTW distance between the normal flight parameter subsequences belonging to the target action in each sortie and the normal flight parameter subsequences belonging to the target action in other sorties is calculated. Select the sortie with the smallest total DTW distance, and use the normal flight parameter subsequence of that sortie as a reference sample for the target flight action; For all historical sorties belonging to the target maneuver's normal flight parameter subsequences, calculate the mean of all DTW distances. and standard deviation The mean DTW distance for each flight maneuver is determined by traversing all historical sorties and the skewness and kurtosis of all DTW distance distributions. and standard deviation , and the skewness and kurtosis of all DTW distance distributions; Based on the skewness and kurtosis of all DTW distance distributions for each flight maneuver, through Calculate the dynamic coefficients for each flight maneuver, where, For the dynamic coefficient of flight maneuvers, , For preset adjustment coefficients, The skewness of the DTW distance distribution for all flight maneuvers. Kuness of the DTW distance distribution for all flight maneuvers; For the flight parameter subsequence to be identified corresponding to the short-time Fourier time-frequency graph of each flight action category in the test set, calculate the DTW distance between the flight parameter subsequence to be identified and the corresponding reference sample. If the DTW distance is within... If the data is outside the specified range, the flight parameter data of the aircraft to be identified is determined to be abnormal.

7. A two-stage flight parameter data anomaly identification device, used to implement the two-stage flight parameter data anomaly identification method according to any one of claims 1-6, characterized in that, include: The flight parameter data acquisition module is used to collect raw flight parameter data of historical sorties and sorties to be identified, and to extract flight parameter subsequences related to flight maneuvers from the raw flight parameter data. The raw flight parameter data is multi-dimensional time-series data including relative pressure altitude, climb rate, radio altimeter indicated altitude, heading angle, pitch angle, airspeed and normal acceleration. The flight parameter data processing module is used to convert the flight parameter subsequence into a short-time Fourier transform (SFT) time-frequency map using SFT, and to label each SFT time-frequency map with flight maneuver categories to form a flight maneuver dataset for the aircraft. The SFT time-frequency maps corresponding to historical sorties in the flight maneuver dataset are divided into a training set, and the SFT time-frequency maps corresponding to the sorties to be identified are divided into a test set. The flight maneuver categories include 180° turn, climb, descent, incorrect control, and maximum angular velocity turn. The deep learning neural network model building module is used to build a deep learning neural network model for recognizing flight action categories. It uses the training set in the flight action dataset to train the model, ensuring that the loss function value is within a preset range, and obtains the trained deep learning neural network model. The flight action classification module is used to take the test set in the flight action dataset as input, and use the trained deep learning neural network model to classify the flight action category of the short-time Fourier time-frequency image in the test set to obtain the classified short-time Fourier time-frequency image. The abnormal data detection module is used to establish an anomaly detection model for single-dimensional raw flight parameter data based on the dynamic time-warping algorithm, and to generate reference samples corresponding to each flight action category using the anomaly detection model; calculate the DTW distance between the flight parameter subsequence to be identified and the corresponding reference sample corresponding to the short-time Fourier time-frequency graph after each classification; if the DTW distance is outside the dynamic preset threshold range, it is determined that the flight parameter data of the aircraft to be identified is abnormal.

8. A computer device comprising a memory (1), a processor (2), and a computer program stored in the memory (1) and executable on the processor (2), characterized in that, When the processor (2) executes the computer program, it implements the two-stage flight parameter data anomaly identification method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that performs the two-stage flight parameter data anomaly identification method according to any one of claims 1 to 7.

Images