A method and system for aircraft stall scene recognition based on deep learning
Through deep learning methods combined with unsupervised and supervised learning, the aircraft stall is dynamically judged, which solves the problem that fixed thresholds in the existing technology cannot adapt to different scenarios, and achieves accurate identification and safety guarantee of aircraft stalls.
Patent Information
- Application Number
- CN202411097459.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-12
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-08-12
AI Technical Summary
The existing deep learning-based aircraft stall prediction and identification methods rely on a fixed stall angle of attack threshold and cannot adapt to the changes in stall angle of attack in different flight scenarios, resulting in aircraft flight safety hazards.
A deep learning-based method is adopted to construct a cluster recognition model through unsupervised learning and a time-sequence prediction model is constructed through supervised learning. The stall prediction model is established based on the two, and whether the aircraft is stalled is dynamically judged and the identification is based on real-time flight data.
It realizes accurate and efficient identification of aircraft stall conditions in different flight scenarios, and is more applicable, ensuring the safety and stability of aircraft flight.
Smart Images

Figure CN118626894B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of aircraft stall, and in particular to an aircraft stall scene recognition method and system based on deep learning. Background Art
[0002] Stall can occur at any stage of flight, and aircraft stall will have an inestimable impact on flight safety. Stall refers to the phenomenon that the aircraft's angle of attack exceeds the angle of approach and cannot maintain normal flight. During the flight of an aircraft, when the relative airflow passes through the upper and lower surfaces of the aircraft's wing, the different flow rates of the airflow on the upper and lower wing surfaces, and then the different pressures, produce a pressure difference. When the pressure generated by the lower wing surface is greater than the lift generated by the upper wing surface, the aircraft gains lift. The fundamental reason for the stall of an aircraft during flight is that the aircraft's angle of attack during flight is greater than the stalling angle, which causes the airflow on the upper surface of the aircraft's wing to separate and produce turbulence. When the pressure generated by the upper wing surface makes the combined force on the wing unable to generate lift to maintain the aircraft's flight, the aircraft enters a stall state.
[0003] The existing stall prediction and identification methods based on deep learning are mainly based on supervised learning methods, and the supervised learning methods require the collected flight parameter data to be labeled. The existing labeling method is based on the judgment method of deterministic stall angle of attack. For example, if it is greater than a given stall angle of attack, it is marked as stalled, and if it is less than a certain stall angle of attack, it is marked as non-stalled. However, since the stall angle of attack has different values in different situations, that is, the stall angle of attack is not a constant, the labeling method of the prior art is different from the actual situation that the stall angle of attack is not a constant. It is not sufficient to rely solely on a certain critical threshold to judge the stall state of the aircraft in various situations, which may easily lead to safety hazards in aircraft flight.
[0004] Therefore, there is an urgent need for an aircraft stall scenario recognition method and system based on deep learning, which can dynamically determine whether the aircraft is stalled according to different flight scenarios, so as to accurately and efficiently identify the aircraft stall situation quickly, with stronger applicability, and ensure the safety and stability of aircraft flight. Summary of the invention
[0005] In order to solve the above technical problems, the present invention provides an aircraft stall scenario recognition method and system based on deep learning, which can dynamically judge whether the aircraft is stalled according to different flight scenarios, so as to accurately and efficiently identify the aircraft stall situation quickly, with stronger applicability, and ensure the safety and stability of aircraft flight.
[0006] The present invention provides an aircraft stall scene recognition method based on deep learning, comprising the following steps:
[0007] S1. Acquire raw data of the target aircraft in different scenarios; wherein the raw data includes flight parameters and aerodynamic shape parameters;
[0008] S2, constructing the first type of data samples and the second type of data samples based on the original data;
[0009] S3, performing unsupervised learning based on the first type of data samples to establish a cluster recognition model;
[0010] S4. Conduct supervised learning based on the second type of data samples to establish a time series prediction model;
[0011] S5. establishing a stall prediction model according to the cluster recognition model and the time series prediction model;
[0012] S6. Inputting the data to be tested of the target aircraft into the stall prediction model to identify the stall scenario of the target aircraft.
[0013] Furthermore, in S2, constructing the first type of data samples based on the original data includes:
[0014] S21a, constructing a first data matrix X according to the original data;
[0015] The first data matrix expression is as follows:
[0016] X=[x ij ];
[0017] Among them, x ij represents the observed value of the i-th data sample on the j-th feature, where the feature is the flight parameters and aerodynamic shape parameters contained in the original data, i=1,2,...,N,j=1,2,...,p, N represents the total number of data samples, and p represents the total number of features;
[0018] S22a, normalize each element in the first data matrix X to obtain a first type of data matrix , as the first type of data samples.
[0019] Furthermore, in S22a, normalizing each element in the first data matrix X includes:
[0020] The Min-Max normalization method is used to normalize each element in the first data matrix X. The normalization formula is as follows:
[0021] ;
[0022] in, represents the observed value of the i-th data sample after normalization on the j-th feature, x ij represents the observed value of the i-th data sample on the j-th feature; xminj represents the minimum value of the observations on the jth feature, x maxj Represents the maximum value of the observations on the jth feature.
[0023] Furthermore, in S2, constructing the second type of data samples based on the original data includes:
[0024] S21b, according to the first type of data matrix Construct a second data matrix Y;
[0025] The second data matrix expression is as follows:
[0026] Y=[y sj ];
[0027] Among them, y sj represents the observed value of the s-th predicted data sample on the j-th feature s=1,2,...,(Nk-n+1),j=1,2,...,p, k represents the time step of the prediction, n represents the historical data step that needs to be recorded for each prediction, (Nk-n+1) represents the total number of predicted data samples, and p represents the total number of features;
[0028] S22b, according to the first type of data matrix Constructing a 3D matrix ;
[0029] S23b, constructing a second type of data matrix XY according to the second data matrix and the three-dimensional matrix as a second type of data sample.
[0030] Furthermore, in S3, unsupervised learning is performed based on the first type of data samples to establish a cluster recognition model including:
[0031] S31, setting the number of clusters to 2, performing clustering calculation on the first type of data samples by using a clustering algorithm, and obtaining two clustering results;
[0032] S32, setting an upper limit value and a lower limit value of the stall angle of attack, and labeling the first type of data samples according to the upper limit value and the lower limit value of the stall angle of attack;
[0033] S33, evaluating the accuracy of the two clustering results obtained by the clustering algorithm according to the labeling results;
[0034] S34, when there is a clustering result with a correct rate greater than a preset correct rate, the clustering result is used as the final classification result;
[0035] When the accuracy of the two clustering results is greater than the preset accuracy, the clustering result with a higher accuracy will be used as the final classification result;
[0036] When the accuracy of the two clustering results is less than or equal to the preset accuracy, the optimization algorithm is used to optimize the clustering algorithm.
[0037] Furthermore, in S4, supervised learning is performed based on the second type of data samples to establish a time series prediction model, including:
[0038] S41, constructing a neural network model;
[0039] Among them, the number of nodes in the input layer is consistent with the number of elements in the three-dimensional matrix in the second type of data samples, and the number of nodes in the input layer is n×p; the number of nodes in the output layer is consistent with the number of elements in the second data matrix in the second type of data samples, and the number of nodes in the output layer is p;
[0040] S42. Divide the second type of data samples into a training set and a test set, train and test the neural network model, and obtain a time series prediction model.
[0041] Furthermore, in S6, inputting the target aircraft's test data into the stall prediction model, and identifying the stall scenario of the target aircraft includes:
[0042] S61, normalizing the data to be measured of the target aircraft; wherein the data to be measured includes the original data of the target aircraft at the current time and (n-1) times before the current time;
[0043] S62, inputting the normalized data to be tested into the time series prediction model of the stall prediction model, and predicting the original data at the kth moment after the current moment;
[0044] S63, inputting the predicted original data at the kth moment into the cluster recognition model of the stall prediction model, and outputting the prediction result;
[0045] S64. Make a judgment based on the prediction result output by the stall prediction model; if the prediction result indicates stall, trigger the alarm mechanism to issue an alarm; if the prediction result indicates normal, do not issue an alarm.
[0046] Furthermore, the flight parameters include: flight angle of attack, flight airspeed, flight attitude, overload, and deceleration rate; wherein the flight attitude includes pitch angle, yaw angle, and roll angle;
[0047] Aerodynamic shape parameters include: flap angle, slat angle, landing gear status, and icing status.
[0048] The present invention also provides an aircraft stall scene recognition system based on deep learning, which is used to execute any of the above-mentioned aircraft stall scene recognition methods based on deep learning, and the system includes the following modules:
[0049] A data acquisition module is used to acquire the original data of the target aircraft in different scenarios; wherein the original data includes flight parameters and aerodynamic shape parameters;
[0050] A data sample construction module, connected to the data acquisition module, is used to construct a first type of data sample and a second type of data sample according to the original data;
[0051] A cluster recognition model building module is connected to the data sample building module and is used to perform unsupervised learning based on the first type of data samples to build a cluster recognition model;
[0052] A time series prediction model building module is connected to the data sample building module and is used to perform supervised learning based on the second type of data samples to build a time series prediction model;
[0053] The stall prediction model establishment module is connected with the cluster recognition model establishment module and the time series prediction model establishment module, and is used to establish the stall prediction model according to the cluster recognition model and the time series prediction model;
[0054] The identification module is connected with the stall prediction model establishment module and is used to input the test data of the target aircraft into the stall prediction model to identify the stall scenario of the target aircraft.
[0055] The embodiments of the present invention have the following technical effects:
[0056] The present invention constructs a cluster recognition model through unsupervised learning, regards the stall angle of attack as a non-constant and labels it intelligently to adapt to the essence of the stall angle of attack being a non-constant, so that the method can be applicable to stall scene recognition under different flight conditions, and then constructs a time series prediction model through supervised learning, and obtains a stall prediction model by combining the cluster recognition model and the time series prediction model. The stall angle of attack of the aircraft is correctly marked according to real-time original flight data, rather than relying solely on a fixed critical stall angle of attack, so as to realize dynamic judgment of whether the aircraft is stalled in different flight scenarios, accurately and efficiently quickly identify the stall situation of the aircraft, and has stronger applicability, thereby ensuring the safe and stable flight of the aircraft. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0058] Figure 1 It is a schematic diagram of the lift-angle of attack curve of a certain passenger aircraft in different configurations;
[0059] Figure 2 is a flow chart of a method for identifying aircraft stall scenarios based on deep learning provided by an embodiment of the present invention;
[0060] Figure 3 It is a schematic diagram comparing the recognition result and the true value of a cluster recognition model provided by an embodiment of the present invention;
[0061] Figure 4 is a confusion matrix diagram of a recognition result of a cluster recognition model provided by an embodiment of the present invention;
[0062] Figure 5 It is a schematic diagram of a training set prediction result of a time series prediction model provided by an embodiment of the present invention;
[0063] Figure 6 It is a schematic diagram of a test set prediction result of a time series prediction model provided by an embodiment of the present invention;
[0064] Figure 7 It is a structural schematic diagram of an aircraft stall scene recognition system based on deep learning provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0065] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention.
[0066] The essence of stall is that the aircraft's angle of attack exceeds the stall angle of attack. In other words, there are two situations in the essence of stall: ① The angle of attack is increased by command or non-command, exceeding the stall angle of attack; ② The stall angle of attack changes (decreases) due to environmental factors and other factors. This is mainly because the aerodynamic characteristics have changed, resulting in a stall even if the angle of attack has not changed, such as when the aircraft is iced. Based on these two points, the following scenarios may occur when the aircraft stalls:
[0067] a. When the stall angle of attack decreases, increase the angle of attack commandedly;
[0068] b. When the load changes due to the need for maneuvers and the airspeed does not increase, the angle of attack is increased uncommandedly to generate sufficient lift;
[0069] c. Changes in flight attitude (such as the appearance of gusts, changes in trajectory during maneuvers that cannot follow the attitude, etc.) cause the angle of attack to increase instantaneously.
[0070] In summary, factors that change the stall angle of attack include: structural deformation, aerodynamic shape differences due to different flap / slat angles, whether the landing gear is retracted or not, icing conditions, etc. Factors that increase the aircraft's angle of attack include: increased overload (especially when the power reserve is insufficient, the angle of attack can only be increased in exchange for lift), reduced speed (especially when the speed is relatively small, special attention should be paid to the deceleration rate), changes in flight attitude, etc.
[0071] To sum up, the stall angle of attack used to judge stall is related to many factors. The value of the stall angle of attack is different in different situations, so the stall angle of attack is not a constant. Figure 1 This is a schematic diagram of the lift-angle of attack curve of a certain passenger aircraft in different configurations, see Figure 1 In the figure, when the angle of attack in the horizontal axis exceeds a certain value, the lift coefficient drops sharply, and the turning point of the sharp drop is the stall angle of attack; the stall angle of attack of this passenger aircraft will be different in different configurations, such as changes in the flap opening angle, whether the landing gear is opened or not, etc. It is not fixed, that is, the stall angle of attack is different for different aircraft configurations.
[0072] In the prior art, the method of labeling the stall angle of attack as a constant and performing stall prediction does not conform to the actual situation. Therefore, this will lead to inaccurate stall judgment, or the established stall identification model can only be applied to limited scenarios. Specifically, the model established based on the stall angle of attack under specific conditions such as landing gear retracted, flaps 0°, and no icing can only effectively identify stall under these specific conditions, and is not applicable to stall prediction under other conditions such as landing gear down, flap angle non-0°, or icing.
[0073] In order to solve the above problems, the present invention proposes an aircraft stall scene recognition method based on deep learning. Figure 2 is a flowchart of a method for identifying aircraft stall scenarios based on deep learning provided by an embodiment of the present invention, see Figure 2 , specifically including:
[0074] S1. Obtain the original data of the target aircraft in different scenarios.
[0075] Specifically, the raw data includes flight parameters and aerodynamic shape parameters.
[0076] Among them, flight parameters may include:
[0077] Flight angle of attack: The flight angle of attack is the angle between the aircraft wing and the relative airflow. When the angle of attack exceeds the stall angle of attack, the aircraft will stall; among them, the stall angle of attack is a non-constant quantity.
[0078] Flying airspeed: The airspeed of an aircraft is an important factor affecting stall. When the airspeed is too low, in order to maintain lift, the aircraft needs to increase the angle of attack, which may cause stall.
[0079] Flight attitude: Changes in flight attitude will affect the angle of airflow passing through the aircraft's wings. When the dive attitude changes rapidly to a level flight attitude, the wing's angle of attack may increase, leading to a stall. The attitude angles of the flight attitude include the pitch angle, yaw angle, and roll angle.
[0080] G-Loads: Increased G-loads may result in an uncommanded increase in the angle of attack, particularly at low speeds and with underpowered engines.
[0081] Deceleration rate: The reduction in speed, especially when the speed is relatively slow, should pay special attention to the deceleration rate because it may affect the change in angle of attack.
[0082] The flight angle of attack, flight airspeed, flight attitude, and overload can be obtained through sensors on the aircraft. The deceleration rate can be obtained through calculation, and the calculation formula is as follows:
[0083] D=(V i+1 -V i ) / t;
[0084] Among them, V i+1 represents the flight speed at the i+1th moment, V i represents the flight speed at the i-th moment, and t represents the time interval between moment i and moment i+1.
[0085] Aerodynamic shape parameters may include:
[0086] Flap angle, slat angle, landing gear status, icing status.
[0087] Among them, the icing conditions include no icing, light icing, moderate icing, and heavy icing.
[0088] The landing gear status includes retracted and extended.
[0089] The flap angle, slat angle and landing gear status are obtained by position sensors on the aircraft.
[0090] S2. Construct the first type of data samples and the second type of data samples based on the original data.
[0091] Specifically, constructing the first type of data samples based on the original data includes:
[0092] S21a, constructing a first data matrix X according to the original data.
[0093] The first data matrix expression is as follows:
[0094] X=[x ij ];
[0095] Among them, x ijIt represents the observed value of the i-th data sample on the j-th feature. The feature is the flight parameters and aerodynamic shape parameters contained in the original data. i=1,2,...,N,j=1,2,...,p, N represents the total number of data samples (i.e. the number of sampling moments), and p represents the total number of features (i.e. the number of flight parameters and aerodynamic shape parameters). Each row of the first data matrix X represents a data sample, and each column represents a feature. The first data matrix X is an N×p matrix.
[0096] S22a, normalize each element in the first data matrix X to obtain a first type of data matrix , as the first type of data samples.
[0097] Specifically, the Min-Max normalization method is used to normalize each element in the first data matrix X. The normalization formula is as follows:
[0098] ;
[0099] in, represents the observed value of the i-th data sample after normalization on the j-th feature, x ij represents the observed value of the i-th data sample on the j-th feature; x minj represents the minimum value of the observations on the jth feature, x maxj Represents the maximum value of the observations on the jth feature.
[0100] After normalization, we get the first type of data matrix , the first type of data matrix As the first type of data sample, the first type of data matrix is a two-dimensional matrix.
[0101] Specifically, constructing the second type of data samples based on the original data includes:
[0102] S21b, according to the first type of data matrix Construct the second data matrix Y.
[0103] The second data matrix expression is as follows:
[0104] Y=[y sj ];
[0105] Among them, y sj Represents the observed value of the s-th predicted data sample on the j-th feature s=1,2,...,(Nk-n+1),j=1,2,...,p, , i=1,2,...,Nk-n+1, k represents the prediction time step, n represents the historical data step that needs to be recorded for each prediction, k and n are set according to the actual situation, (Nk-n+1) represents the total number of prediction data samples, and p represents the total number of features.
[0106] For example, if the sampling time interval is t=1s, k=5, and n=20, it means that the aircraft parameters of the 25th second are predicted based on the 20 data generated in 20s. The values of k and n are determined according to the aircraft type. If the aircraft has strong maneuverability and agility, n is larger and k is smaller. On the contrary, if the aircraft has low maneuverability and agility, n is smaller and k is larger. k represents the time reserved for the pilot or the aircraft controller to recover from the impending stall state.
[0107] S22b, according to the first type of data matrix Constructing a 3D matrix .
[0108] The three-dimensional matrix expression is as follows:
[0109] ;
[0110] Among them, the three-dimensional matrix The size of is (Nk-n+1)×n×p, represents the first n×p matrix; represents the second n×p matrix; Represents the (Nk-n+1)th n×p matrix.
[0111] S23b, constructing a second type of data matrix XY according to the second data matrix and the three-dimensional matrix as a second type of data sample.
[0112] Specifically, the size of the second type data matrix XY is (Nk-n+1)×(n+1)×p, the number of samples of the second type data matrix XY is (Nk-n+1), and the rth sample of the second type data matrix XY is as follows, r=1,2,…,Nk-n+1:
[0113] ;
[0114] Will The input sample of the rth sample of the second type of data matrix XY is called It is called the output sample of the rth sample of the second type data matrix XY.
[0115] S3. Perform unsupervised learning based on the first type of data samples to establish a cluster recognition model.
[0116] S31. Set the number of clusters to 2, perform clustering calculation on the first type of data samples through a clustering algorithm, and obtain two clustering results.
[0117] The selection of clustering algorithm needs to meet two conditions:
[0118] ①The number of clusters needs to be specified before the clustering algorithm can be implemented;
[0119] ②The clustering algorithm needs to have hyperparameters, that is, by adjusting the hyperparameters, different clustering results can be obtained.
[0120] The clustering results are expressed as L ia Indicates that L ia =0 or 1, i=1,2,…N, 0 means normal, 1 means stall, and a means the classification result obtained by using the clustering algorithm. Since the clustering algorithm is based on unlabeled data, even if the clustering results are obtained, it is not possible to directly determine which category in the clustering results belongs to stall and which category belongs to normal. Therefore, using the clustering method, the stall problem can obtain two types of clustering results: clustering results of method 1 and clustering results of method 2.
[0121] For example, the clustering algorithm can be K-means algorithm, and the hyperparameter of K-means algorithm is the initial position of cluster center. For the problem of stall state recognition, it is essentially a two-classification problem: stall and normal. When using K-means algorithm to classify the first type of data samples, it is necessary to initialize two initial positions: C 1 =(c 11 ,c 12 ,…,c 1j ,…,c 1p ), C 2 =(c 21 ,c 22 ,…,c 2j ,…,c 2p ), where c 1j and c 2j is a random number uniformly distributed on [0, 1]. 1j and c 2j Indicates the initial position C 1 and C 2 The possible values of the jth feature in . Perform clustering calculations using the K-means algorithm and record the center position after the algorithm is completed. and , according to the first type of data matrix The clustering result is determined by the Euclidean distance between each row of data and the two center locations.
[0122] Method 1 Clustering results: If the distance If the distance If they are close, the clustering result is marked as 1. If the distances between the two are the same, they are randomly marked as 0 or 1.
[0123] Method 2 Clustering results: If the distance If the clustering result is close, it is marked as 1. If they are close, the clustering result is marked as 0. If the distances between the two are the same, they are randomly marked as 0 or 1.
[0124] S32. Setting an upper limit value and a lower limit value of the stall angle of attack, and labeling the first type of data samples according to the upper limit value and the lower limit value of the stall angle of attack.
[0125] Specifically, although the stall angle of attack takes different values in different situations and is not a constant, its range of variation must be within a reasonable range. Generally speaking, the upper and lower limits of the stall angle of attack of the target aircraft can be set according to the different aircraft types.
[0126] According to the upper and lower limits of the stall angle of attack, some data in the first category of data samples can be labeled:
[0127] If the angle of attack in the i-th row sample in the first data matrix X exceeds the upper limit of the stall angle of attack, the i-th sample of the first category of data samples is labeled 1, indicating stall; if the angle of attack in the i-th row sample in the first data matrix X is less than the lower limit of the stall angle of attack, the i-th sample of the first category of data samples is labeled 0, indicating normal. Therefore, some data in the first category of data samples are labeled, and some data are not labeled. The unlabeled data is marked with -1, and -1 indicates that no identification is performed. This labeling method is a threshold-based labeling method, and the result of labeling based on the threshold labeling method is expressed as L ib , L ib =0 or 1 or -1, i=1,2,…N, b represents the classification result obtained by using the threshold judgment method, L ib =1 or 0 is represented by N', N'<N, that is, N' represents the number of samples that can be labeled for the first type of data samples using the threshold judgment method.
[0128] For example, taking the Boeing 737 passenger plane as an example, the stall angle of attack is different under different flight conditions, but no matter what configuration or state it is in, its stall angle of attack ranges from 15° to 20°. Therefore, the upper limit of the stall angle of attack of the Boeing 737 is set to 20°, and the lower limit of the stall angle of attack is set to 15°. For the Boeing 737 passenger plane, if the characteristic value of the angle of attack at a certain moment is greater than 20°, then the label value is set to 1 for the characteristic parameter at that moment. If the characteristic value of the angle of attack at a certain moment is less than 15°, the label value is set to 0. If the characteristic value of the angle of attack at a certain moment is between [15°, 20°], no setting is performed.
[0129] S33. Evaluate the accuracy of the two clustering results obtained by the clustering algorithm according to the labeling results.
[0130] Specifically, the accuracy evaluation formula is as follows:
[0131] ;
[0132] Among them, P represents the accuracy, and the subscript u represents L ib =0 or 1 sample number, u=i 1 ,i 2 ,…i N’ ,i 1 Indicates the first L in N samples ib The sample number is equal to 0 or 1, i 2 Indicates the second L in N samples ib The sample number equal to 0 or 1, ..., i N’ Indicates the N'th L in N samples ib The sample number equal to 0 or 1, L ub represents the data samples whose classification results are 0 or 1 based on the threshold judgment method, N' represents the total number of data samples whose classification results are 0 or 1 that can be labeled using the threshold judgment method, and L ua Represents the clustering result L obtained based on the clustering algorithm ia In, with L ub The data samples corresponding to the sample numbers in are substituted into the evaluation formulas for the clustering results of method 1 and method 2, respectively, to obtain the accuracy of the two clustering results.
[0133] Furthermore, L ia , L ib The range is greater than L ua , L ub , L ia , L ib All data samples will be labeled. Since the threshold judgment method will only label some data samples as 0 or 1, the clustering algorithm will classify all data samples as 0 or 1. Therefore, L ia , L ib Some data samples cannot be compared, so we need to select only some data samples with label results of 0 or 1 from all data samples. ua , L ub Calculate the accuracy.
[0134] For example, suppose there are 10 data samples (x 1 、x 2 、x 3 、x 4 、x 5 、x6 、x 7 、x 8 、x 9 、x 10 ), then the clustering algorithm must label all 10 data samples (the label result is L 1a , L 2a , L 3a , L 4a , L 5a , L 6a , L 7a , L 8a , L 9a , L 10a ), while the threshold judgment method only labels 6 of them (for example, only x 1 、x 2 、x 3 、x 6 、x 7 、x 10 Label, the label result is L 1b , L 2b , L 3b , L 6b , L 7b , L 10b ), then the 6 L involved in the accuracy calculation are ia Value (L 1a , L 2a , L 3a , L 6a , L 7a , L 10a ) and 6 L ib Value (L 1b , L 2b , L 3b , L 6b , L 7b , L 10b ), these 6 L ia The value is recorded as L ua , these 6 L ib The value is recorded as L ub .
[0135] S34, when there is a clustering result with a correct rate greater than a preset correct rate, the clustering result is used as the final classification result;
[0136] When the accuracy of the two clustering results is greater than the preset accuracy, the clustering result with a higher accuracy will be used as the final classification result;
[0137] When the accuracy of the two clustering results is less than or equal to the preset accuracy, the optimization algorithm is used to optimize the clustering algorithm.
[0138] Furthermore, the clustering algorithm is optimized using the optimization algorithm, specifically including:
[0139] The selection requirement of the optimization algorithm is: it should be able to handle the iterative generation problem of hyperparameters in the clustering algorithm, that is, the optimization algorithm should have the ability to generate a new generation of hyperparameters that make the objective function better from the previous generation of hyperparameters.
[0140] For example, Figure 3 is a schematic diagram of a comparison between a recognition result of a cluster recognition model provided by an embodiment of the present invention and a true value, Figure 4 is a schematic diagram of a confusion matrix of a recognition result of a cluster recognition model provided by an embodiment of the present invention, see Figure 3-Figure 4 . Figure 3 The comparison shown is to compare the recognition results obtained by using the trained clustering model with the true value of the sample. Figure 3 In the figure, the vertical axis is the stall mark, 0 means normal, 1 means stall, and the horizontal axis is the sample number, with a total of 15,000 samples. Figure 3 The higher the degree of agreement between the predicted value and the true value, the higher the accuracy of the clustering model. Figure 3 The "true value" here is the label value 0 or 1 obtained by the threshold method, while the "recognition value" refers to the result obtained by the clustering method. Figure 4 It shows that, using the clustering model, a total of 7375+7465=14830 samples can be correctly identified as stalled, 82 samples are mistakenly considered normal, and 78 samples are mistakenly considered stalled. For the K-means algorithm, if the accuracy of the clustering results of method 1 and method 2 is not greater than the preset accuracy, for example, 90%, the first type of data samples can be used to generate two new position parameters C using the genetic algorithm in the optimization algorithm. 1-新 =(c 11-新 ,c 12-新 ,…,c 1j-新 ,…,c 1p-新 ) and C 2-新 =(c 21-新 ,c 22-新 ,…,c 2j-新 ,…,c 2p-新 ).
[0141] The basic steps for optimizing the accuracy of the K-means algorithm using a genetic algorithm are:
[0142] I) Set the population size of the genetic algorithm to popsize;
[0143] Initialize position C for each individual 1 and C 2 ;
[0144] Using the method described in step S31, for position C 1 and C 2 Generates a population of size popsize.
[0145] For each individual in the population, a value is randomly drawn from the uniform distribution [0,1] as C 1 and C 2 The initial value of .
[0146] II) Taking the accuracy as the objective function, using the first type of data samples, and using the method described in step S33 for a population of size popsize, calculate the accuracy corresponding to each individual in the population.
[0147] III) Based on the accuracy, use the selection algorithm, crossover algorithm and mutation algorithm to generate a new population of size popsize; the selection algorithm can select the roulette wheel algorithm, the crossover algorithm can select the single-point crossover algorithm, and the mutation algorithm can select the single-point mutation algorithm.
[0148] The roulette wheel selection method ensures that the principle of survival of the fittest is reflected in the genetic algorithm, that is, individuals with high accuracy are more likely to be retained and passed on to the next generation. Through continuous iteration of selection, crossover and mutation processes, the genetic algorithm can continuously generate individuals with higher accuracy, thereby achieving the goal of improving the overall accuracy.
[0149] Repeat steps II) and III) until the optimization end condition is met. The optimization end condition includes:
[0150] The first method is to set a preset accuracy rate, that is, when the objective function meets the preset accuracy rate, such as 95%, the optimization can be ended, otherwise it returns to execute step II) and step III).
[0151] The second method is to give a maximum number of iterations. When the maximum number of iterations is reached, the optimization ends, otherwise the process returns to step II) and step III).
[0152] S4. Perform supervised learning based on the second type of data samples to establish a time series prediction model.
[0153] S41. Construct a neural network model.
[0154] Specifically, CNN, LSTM, Bi-LSTM, etc. can be selected as the neural network model of this solution; the neural network model includes an input layer, an intermediate layer, an output layer and hyperparameters.
[0155] The hyperparameters of the neural network model, the number of intermediate layers, and the number of nodes in each intermediate layer are set according to actual conditions.
[0156] The number of nodes in the input layer is consistent with the number of elements in the three-dimensional matrix in the second type of data sample (that is, the number of each element in the input sample of the second type of data matrix XY), and the number of nodes in the input layer is n×p;
[0157] The number of nodes in the output layer is consistent with the number of elements of the second data matrix in the second type of data samples, that is, the number of each element in the output sample of the second type of data matrix XY), and the number of nodes in the output layer is p.
[0158] S42. Divide the second type of data samples into a training set and a test set, train and test the neural network model, and obtain a time series prediction model.
[0159] Specifically, Figure 5 is a schematic diagram of a training set prediction result of a time series prediction model provided by an embodiment of the present invention, Figure 6 is a schematic diagram of the prediction results of a test set of a time series prediction model provided by an embodiment of the present invention, see Figure 5 and Figure 6 The time series prediction model is constructed using the long short-term memory network model (LSTM model), where k = 5 and n = 20. The second type of data samples are divided into a training set (70%-80%) and a test set (20%-30%), and the neural network model is trained and tested. By adjusting the hyperparameters, the neural network model with the best test effect is used as the time series prediction model.
[0160] S5. Establish a stall prediction model based on the cluster recognition model and the time series prediction model.
[0161] S6. Inputting the data to be tested of the target aircraft into the stall prediction model to identify the stall scenario of the target aircraft.
[0162] S61. Normalize the data to be measured of the target aircraft.
[0163] The data to be tested include the original data of the target aircraft at the current time and the (n-1) time points before the current time. These original data are normalized to ensure the consistency and comparability of the data.
[0164] S62: input the normalized data to be tested into the time series prediction model of the stall prediction model, and predict the original data at the kth moment after the current moment.
[0165] S63, inputting the predicted original data at the kth moment into the cluster recognition model of the stall prediction model, and outputting the prediction result.
[0166] S64. Making a judgment according to the prediction result output by the stall prediction model.
[0167] Specifically, if the prediction result is 1, it means that the device or system will stall at the kth moment in the future, and the alarm mechanism is triggered to alarm; if the prediction result is 0, it means that the device or system will operate normally at the kth moment in the future, and no alarm is triggered.
[0168] The present invention constructs a cluster recognition model through unsupervised learning, regards the stall angle of attack as a non-constant and labels it intelligently to adapt to the essence of the stall angle of attack being a non-constant, so that the method can be applicable to stall scene recognition under different flight conditions, and then constructs a time series prediction model through supervised learning, and obtains a stall prediction model by combining the cluster recognition model and the time series prediction model. The stall angle of attack of the aircraft is correctly marked according to real-time original flight data, rather than relying solely on a fixed critical stall angle of attack, so as to realize dynamic judgment of whether the aircraft is stalled in different flight scenarios, accurately and efficiently quickly identify the stall situation of the aircraft, and has stronger applicability, thereby ensuring the safe and stable flight of the aircraft.
[0169] Figure 7 is a schematic diagram of the structure of a system for recognizing aircraft stall scenes based on deep learning provided by an embodiment of the present invention, and the system is used to execute the method for recognizing aircraft stall scenes based on deep learning described in the above embodiment, such as Figure 7 As shown, the system includes the following modules:
[0170] A data acquisition module is used to acquire the original data of the target aircraft in different scenarios; wherein the original data includes flight parameters and aerodynamic shape parameters;
[0171] A data sample construction module, connected to the data acquisition module, is used to construct a first type of data sample and a second type of data sample according to the original data;
[0172] A cluster recognition model building module is connected to the data sample building module and is used to perform unsupervised learning based on the first type of data samples to build a cluster recognition model;
[0173] A time series prediction model building module is connected to the data sample building module and is used to perform supervised learning based on the second type of data samples to build a time series prediction model;
[0174] The stall prediction model establishment module is connected with the cluster recognition model establishment module and the time series prediction model establishment module, and is used to establish the stall prediction model according to the cluster recognition model and the time series prediction model;
[0175] The identification module is connected with the stall prediction model establishment module and is used to input the test data of the target aircraft into the stall prediction model to identify the stall scenario of the target aircraft.
[0176] It should be noted that the terms used in the present invention are only for describing specific embodiments, rather than limiting the scope of the present application. As shown in the present specification, unless the context clearly indicates an exception, the words "one", "a", "a kind of" and / or "the" do not specifically refer to the singular, but may also include the plural. The terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method or device. In the absence of more restrictions, the elements defined by the sentence "include one..." do not exclude the presence of other identical elements in the process, method or device including the elements.
[0177] It should also be noted that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. Unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", etc. should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be a connection between the two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.
Claims
1. A method for aircraft stall scene recognition based on deep learning, It is characterized in that The steps include: S1. Acquire raw data of the target aircraft in different scenarios; wherein the raw data includes flight parameters and aerodynamic shape parameters; S2. constructing a first type of data sample and a second type of data sample according to the original data; S3. Perform unsupervised learning based on the first type of data samples to establish a cluster recognition model; The S3 specifically includes: S31, setting the number of clusters to 2, performing clustering calculation on the first type of data samples by using a clustering algorithm, and obtaining two clustering results; S32, setting an upper limit value and a lower limit value of the stall angle of attack, and labeling the first type of data samples according to the upper limit value and the lower limit value of the stall angle of attack; S33, evaluating the accuracy of the two clustering results obtained by the clustering algorithm according to the labeling results; S34, when there is a clustering result with a correct rate greater than a preset correct rate, the clustering result is used as the final classification result; When the accuracy of the two clustering results is greater than the preset accuracy, the clustering result with a higher accuracy will be used as the final classification result; When the accuracy of the two clustering results is less than or equal to the preset accuracy, the clustering algorithm is optimized using an optimization algorithm; S4. Perform supervised learning based on the second type of data samples to establish a time series prediction model; S5. Establishing a stall prediction model according to the cluster recognition model and the time series prediction model; S6, inputting the data to be measured of the target aircraft into the stall prediction model to identify the stall scenario of the target aircraft; The S6 specifically includes: S61, normalizing the data to be measured of the target aircraft; wherein the data to be measured includes the original data of the target aircraft at the current time and (n-1) times before the current time; S62, inputting the normalized test data into the time series prediction model of the stall prediction model, and predicting the original data at the kth moment after the current moment; S63, inputting the predicted original data at the kth moment into the cluster recognition model of the stall prediction model, and outputting the prediction result; S64. Making a judgment according to the prediction result output by the stall prediction model; if the prediction result indicates stall, triggering an alarm mechanism to give an alarm; if the prediction result indicates normal, no alarm is given.
2. According to claim 1, a method for recognizing aircraft stall scenarios based on deep learning, It is characterized in that In S2, constructing a first type of data sample according to the original data includes: S21a, constructing a first data matrix X according to the original data; The first data matrix expression is as follows: X=[x ij ]; Among them, x ij represents the observed value of the i-th data sample on the j-th feature, where the feature is the flight parameters and aerodynamic shape parameters contained in the original data, i=1,2,...,N,j=1,2,...,p, N represents the total number of data samples, and p represents the total number of features; S22a, normalize each element in the first data matrix X to obtain a first type of data matrix , as the first type of data samples.
3. According to claim 2, a method for recognizing aircraft stall scenarios based on deep learning, It is characterized in that In S22a, normalizing each element in the first data matrix X includes: The Min-Max normalization method is used to normalize each element in the first data matrix X. The normalization formula is as follows: ; in, represents the observed value of the i-th data sample after normalization on the j-th feature, x ij represents the observed value of the i-th data sample on the j-th feature; x minj represents the minimum value of the observations on the jth feature, x maxj Represents the maximum value of the observations on the jth feature.
4. According to claim 2, a method for recognizing aircraft stall scenarios based on deep learning, It is characterized in that In S2, constructing the second type of data samples according to the original data includes: S21b, according to the first type of data matrix Construct a second data matrix Y; The second data matrix expression is as follows: Y=[y sj ]; Among them, y sj represents the observed value of the s-th predicted data sample on the j-th feature s=1,2,...,(Nk-n+1),j=1,2,...,p, k represents the time step of the prediction, n represents the historical data step that needs to be recorded for each prediction, (Nk-n+1) represents the total number of predicted data samples, and p represents the total number of features; S22b, according to the first type of data matrix Constructing a 3D matrix ; S23b, constructing a second type of data matrix XY according to the second data matrix and the three-dimensional matrix as a second type of data sample.
5. According to claim 4, a method for recognizing aircraft stall scenarios based on deep learning, It is characterized in that In S4, the supervised learning is performed based on the second type of data samples to establish a time series prediction model, which includes: S41, constructing a neural network model; Among them, the number of nodes in the input layer is consistent with the number of elements in the three-dimensional matrix in the second type of data samples, and the number of nodes in the input layer is n×p; the number of nodes in the output layer is consistent with the number of elements in the second data matrix in the second type of data samples, and the number of nodes in the output layer is p; S42. Divide the second category of data samples into a training set and a test set, train and test the neural network model, and obtain a time series prediction model.
6. The aircraft stall scene recognition method based on deep learning according to claim 1, It is characterized in that In S1, the flight parameters include: flight angle of attack, flight airspeed, flight attitude, overload, and deceleration rate; wherein the flight attitude includes pitch angle, yaw angle, and roll angle; The aerodynamic shape parameters include: flap angle, slat angle, landing gear status, and icing status.
7. An aircraft stall scene recognition system based on deep learning, used to execute the aircraft stall scene recognition method based on deep learning according to any one of claims 1 to 6 above, It is characterized in that The system includes the following modules: A data acquisition module, used to acquire raw data of the target aircraft in different scenarios; wherein the raw data includes flight parameters and aerodynamic shape parameters; A data sample construction module, connected to the data acquisition module, for constructing a first type of data sample and a second type of data sample according to the original data; A clustering recognition model establishment module is connected to the data sample construction module, and is used to perform unsupervised learning based on the first type of data samples to establish a clustering recognition model; specifically comprising: setting the number of clusters to 2, performing clustering calculations on the first type of data samples through a clustering algorithm, and obtaining two clustering results; setting an upper limit value and a lower limit value of the stall angle of attack, and labeling the first type of data samples according to the upper limit value and the lower limit value of the stall angle of attack; evaluating the accuracy of the two clustering results obtained by the clustering algorithm according to the labeling results; when there is a clustering result with an accuracy greater than a preset accuracy, the clustering result is used as the final classification result; when the accuracy of the two clustering results is greater than the preset accuracy, the clustering result with a higher accuracy is used as the final classification result; when the accuracy of the two clustering results is less than or equal to the preset accuracy, the clustering algorithm is optimized using an optimization algorithm; A time series prediction model building module, connected to the data sample building module, for conducting supervised learning based on the second type of data samples to build a time series prediction model; A stall prediction model establishing module, connected to the cluster recognition model establishing module and the time series prediction model establishing module, and used to establish a stall prediction model according to the cluster recognition model and the time series prediction model; The identification module is connected to the stall prediction model establishment module and is used for inputting the test data of the target aircraft into the stall prediction model to identify the stall scenario of the target aircraft; specifically comprising: normalizing the test data of the target aircraft; wherein the test data includes the original data of the target aircraft at the current moment and (n-1) moments before the current moment; inputting the normalized test data into the time series prediction model of the stall prediction model, and predicting the original data at the kth moment after the current moment; inputting the predicted original data at the kth moment into the cluster recognition model of the stall prediction model, and outputting the prediction result; making a judgment according to the prediction result output by the stall prediction model; if the prediction result indicates stall, triggering an alarm mechanism to alarm; if the prediction result indicates normal, no alarm is issued.
Citation Information
Patent Citations
Flight action intelligent identification method
CN114298183A
Over-stall prediction method based on multi-task Transformer
CN116484497A
Unmanned aerial vehicle intelligent air target grouping method based on unsupervised comparative learning
CN117493915A
Gas compressor stall prediction method, device and equipment based on LSTM neural network and medium
CN117909930A
Flight safety risk assessment method and device, electronic equipment and storage medium
CN117972336A
