A multi-feature flight maneuver recognition method
Through the multi-feature flight maneuver action recognition method, the convolutional neural network model is used to extract the characteristics of flight parameter data, which solves the problem of low recognition accuracy in the prior art and achieves more efficient flight action recognition.
Patent Information
- Application Number
- CN202410320653.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-20
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-03-20
AI Technical Summary
The method of identifying aircraft flight actions in the prior art requires a large amount of expert action information, which leads to high time consumption and low recognition accuracy of complex maneuver actions.
The multi-feature flight maneuver action recognition method is adopted to obtain and preprocess flight parameter data, build a convolutional neural network model, extract feature data, and update the model through loss calculation to improve recognition accuracy and efficiency.
It improves the recognition accuracy and efficiency of flight actions, and enhances the recognition accuracy and generalization of flight parameter data.
Smart Images

Figure CN118364231B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of aircraft motion recognition, and in particular to a multi-feature flight maneuver recognition method. Background Art
[0002] The evaluation of flight training quality is achieved by evaluating the completion and quality of maneuvers in flight training subjects. The segmented identification of maneuvers is achieved by analyzing the time series data recorded during the flight training process (referred to as flight parameter data).
[0003] The methods disclosed in the prior art for identifying the flight actions of aircraft mainly include manual discrimination methods and probabilistic model reasoning; the manual discrimination method is to construct an expert action library through prior domain expert knowledge, and segment the flight trajectory and correspond to expert actions in combination with the recorded flight parameter data; while the probabilistic model reasoning is to construct a Bayesian network by selecting flight parameters as nodes of the Bayesian network, calculate the conditional probability and state transition probability through prior expert information, continuously iterate and update the probability value, and finally select the one with the largest inference probability as the predicted action.
[0004] However, the manual discrimination methods and probability model reasoning disclosed in the prior art both require a large amount of expert action information for preliminary preparation, which consumes a lot of time and has low recognition accuracy for complex maneuvers. Summary of the invention
[0005] The embodiment of the present application provides a multi-feature flight maneuver recognition method, which can improve the accuracy of identifying flight actions to a certain extent.
[0006] The present application provides a multi-feature flight maneuver recognition method, the method comprising:
[0007] Acquire and preprocess the flight parameter data, and divide it into a training data set and a test data set;
[0008] Construct a convolutional neural network model, and construct multiple convolution output channels corresponding to different dimensions after the convolutional neural network model;
[0009] Load the flight parameter data of the training data set into the convolutional neural network model, output the prediction feature graphs corresponding to different dimensions, decode and extract the feature data in the corresponding prediction feature graphs;
[0010] Load the flight parameter data of the training data set, decode and convert the flight parameter data into an expression consistent with the form of the extracted feature data; when the number of training times reaches the preset number, calculate the loss and update the convolutional neural network model;
[0011] The flight parameter data in the test data set is predicted using the updated convolutional neural network model.
[0012] In some embodiments, obtaining and preprocessing the flight parameter data includes:
[0013] Collecting flight parameter data of the aircraft; wherein the flight parameter data includes: flight time, flight altitude, aircraft flight speed and angle of attack;
[0014] Integrate the acquired flight parameter data of the aircraft and perform data smoothing;
[0015] Extract the time range corresponding to the maneuver in the flight data, and add category labels to the corresponding maneuvers in the flight data.
[0016] In some embodiments, the obtaining and preprocessing of the flight parameter data further comprises:
[0017] When the amount of flight parameter data obtained is small, the flight altitude change rate, track quaternion change rate, aircraft attitude angle, aircraft attitude angle change rate, and additional state quantities of angular velocity rotation under the body axis system are obtained through differential methods according to the flight time, flight altitude, and track quaternion. At the same time, offsets are added to the flight altitude and speed to obtain sample data of the same trajectory at other spatial positions, and the obtained data is added to the flight parameter data to expand the data volume.
[0018] In some embodiments, constructing a convolutional neural network model includes:
[0019] Clustering is performed according to the duration corresponding to all maneuvers in the flight parameter data, and the widths of multiple feature map windows corresponding to different dimensions are obtained respectively;
[0020] A convolutional neural network model is constructed by combining one-dimensional convolution with a residual network, and the convolutional neural network model corresponds to the common parts of feature maps of multiple different dimensions.
[0021] In some embodiments, constructing a plurality of convolution output channels corresponding to different dimensions after the convolutional neural network model includes:
[0022] The size of each window width is obtained, and the dimension of the convolution output channel to be constructed is obtained according to the size of each window; the convolution output channel is constructed after the convolution neural network model according to the dimension of the convolution output channel to be constructed.
[0023] In some embodiments, the constructed convolution output channel includes:
[0024] At least seven convolutional layers of the neural network, of which the first five layers are used to extract the features of the flight parameter data of the training data set, and the last two layers are used to correspond to the prediction outputs of different dimensions in the prediction feature map.
[0025] In some embodiments, in the decoding and extraction of feature data in the corresponding predicted feature map, the feature data includes: a priori box center point adjustment parameter, a priori box width adjustment parameter, action existence confidence and action type confidence.
[0026] In some embodiments, decoding and converting the flight parameter data into an expression consistent with the form of the extracted feature data includes:
[0027] Convert the intermediate time information and duration information in the flight parameter data corresponding to the feature data into the corresponding feature layer, and obtain the center point and width of the real frame represented in the feature layer;
[0028] Calculate the intersection-over-union ratio between the true frame and the prior frames of each width obtained by clustering, output the prior frame with the highest overlap, and calculate the offset between the center point of the corresponding prior frame and the true frame and the width offset;
[0029] The flight parameter data with action confidence and action type confidence are encoded to obtain an expression format consistent with the form of extracted feature data.
[0030] In some embodiments, performing loss calculation and updating the convolutional neural network model includes:
[0031] The offset loss is adjusted by calculating the center point through cross entropy;
[0032] The width adjustment offset loss is calculated by the mean square error;
[0033] The confidence loss of action existence and action category confidence is calculated through cross entropy loss;
[0034] The center point offset loss, width adjustment loss, and loss calculation action existence confidence and action category confidence loss are combined to update the parameters of the constructed convolutional neural network model.
[0035] In some embodiments, predicting the flight parameter data in the test data set by using the updated convolutional neural network model includes:
[0036] The flight parameter data corresponding to the test data set is input into the updated convolutional neural network model to obtain prediction outputs of three different feature dimensions;
[0037] The obtained prediction output is compared with the flight parameter data corresponding to the training data set to calculate the intersection and union ratio, and the one with the largest intersection and union ratio is selected as the final prediction output of the model.
[0038] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:
[0039] The embodiment of the present application first obtains flight parameter data, pre-processes the flight parameter data, and then divides the flight parameter data into a training data set and a test data set for training a convolutional neural network model. Then, feature information is extracted from the flight parameter data corresponding to the training data set by using time convolution windows of different step lengths, and the convolutional neural network model is trained using the acquired multi-feature graph, thereby achieving the purpose of improving the efficiency and accuracy of flight action recognition. At the same time, during the training process, the convolutional neural network model is adjusted by performing loss calculation and calculating the loss parameters obtained by calculation, thereby improving the accuracy and generalization of flight parameter data recognition to a certain extent, thereby further achieving the purpose of improving the accuracy and efficiency of flight action recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments of the present application will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0041] Figure 1 A flowchart of a multi-feature flight maneuver identification method provided in an embodiment of the present application;
[0042] Figure 2 A flowchart of the implementation process of the multi-feature flight maneuver recognition method provided in an embodiment of the present application;
[0043] Figure 3 A schematic diagram of the structure of a convolutional neural network model established for the multi-feature flight maneuver recognition method provided in an embodiment of the present application;
[0044] Figure 4 A schematic diagram of the structure of a down-sampling ResConv network of a convolutional neural network model established for the multi-feature flight maneuver recognition method provided in an embodiment of the present application;
[0045] Figure 5 A schematic diagram of the structure of the convolution output channel of the convolutional neural network model established for the multi-feature flight maneuver recognition method provided in an embodiment of the present application. DETAILED DESCRIPTION
[0046] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0047] Reference Figure 1 and Figure 2 , Figure 1 A flowchart of a multi-feature flight maneuver identification method provided in an embodiment of the present application; Figure 2 A flowchart of the implementation process of the multi-feature flight maneuver recognition method provided in an embodiment of the present application.
[0048] An embodiment of the present application provides a multi-feature flight maneuver recognition method. The method provided by the present application includes: step 101, step 102, step 103, step 104 and step 105.
[0049] Step 101, obtain and pre-process the flight parameter data, and divide it into a training data set and a test data set.
[0050] Exemplarily, flight parameter data is acquired through relevant airborne equipment carried by the aircraft, and then the flight parameter data is smoothed and enhanced to increase the sample size of the flight parameter data. The data is then divided into training samples and test samples in a ratio of 9:1 for training and testing of the convolutional neural network recognition model.
[0051] Step 102, construct a convolutional neural network model, and construct multiple convolution output channels corresponding to different dimensions after the convolutional neural network model.
[0052] Step 103, load the flight parameter data of the training data set into the convolutional neural network model, output the prediction feature graphs corresponding to different dimensions, decode and extract the feature data in the corresponding prediction feature graphs.
[0053] Step 104, load the flight parameter data of the training data set, decode and convert the flight parameter data into an expression consistent with the form of the extracted feature data; when the number of training times reaches the preset number, calculate the loss and update the convolutional neural network model.
[0054] Step 105, predicting the flight parameter data in the test data set through the updated convolutional neural network model.
[0055] By first obtaining the flight parameter data and preprocessing the flight parameter data, the flight parameter data is divided into a training data set and a test data set for training the convolutional neural network model. Then, feature information is extracted from the flight parameter data corresponding to the training data set by using time convolution windows of different step lengths, and the convolutional neural network model is trained using the obtained multi-feature graph, thereby achieving the purpose of improving the efficiency and accuracy of flight action recognition. At the same time, during the training process, the loss calculation is performed and the convolutional neural network model is adjusted according to the calculated loss parameters, which improves the accuracy and generalization of the recognition of flight parameter data to a certain extent, further achieving the purpose of improving the accuracy and efficiency of recognizing flight actions.
[0056] In some embodiments, when acquiring and preprocessing flight parameter data, the method provided in the present application includes: collecting flight parameter data of an aircraft; integrating the acquired flight parameter data of the aircraft and performing data smoothing on it; extracting the time range corresponding to the maneuvering action in the flight parameter data, and adding category labels to the corresponding maneuvering actions in the flight parameter data; wherein the flight parameter data to be collected in the present application include: flight time, flight altitude, aircraft flight airspeed and angle of attack.
[0057] Table 1 Comparison table of maneuvers and number labels
[0058] Maneuver name Numbered labels Inverted flight recovery 0 pound 1 roll 2 Lazy "8" 3 Barrel Roll 4 Half roll inverted 5 Half-pound bucket flip 6 Next "8" 7 Inclined bucket 8 Sharp rising turn 9 Clean stall 10 High nose recovery 11 Oblique tumble + downward "8" 12 Maneuvering Stall 13 Route Stall 14 Tilting + flipping 15
[0059] Exemplarily, flight parameter data is acquired through relevant airborne equipment carried by the aircraft, and the flight parameter data recorded by the airborne equipment is loaded and integrated, and data smoothing processing is performed on the flight parameter data. By performing data smoothing processing on the flight parameter data, it is possible to eliminate wild values in the recorded data and eliminate data glitches based on the average of relevant data before and after the recording time point.
[0060] By inputting the dimension used to represent the time sample of the same flight parameter data sequence, extracting the maneuvering action part time in the flight parameter data, reducing the maneuvering time range that is too long, and increasing the maneuvering time range that is too short, the flight parameter data is uniformly divided into a sequence of 3008 frames in length in the embodiment of the present application, and a category label is added to the segment of flight parameter data according to the preset maneuvering action number. The information contained in the category label includes the normalized information of the middle moment of the maneuvering action in the whole sequence, the normalized information of the duration of the maneuvering action in the whole sequence, and the type number of the maneuvering action.
[0061] In some embodiments, in order to reduce the probability that the applicability of the model acquired through training is low due to the small amount of flight parameter data acquired, when acquiring and preprocessing the flight parameter data, the method provided in the present application also includes: when the amount of flight parameter data acquired is small, based on the flight time, flight altitude, and track quaternion, the flight altitude change rate, track quaternion change rate, aircraft attitude angle, aircraft attitude angle change rate, and additional state quantities of the angular velocity of rotation under the body axis system are obtained by differential methods, and the relevant flight parameter data are expanded to 25 dimensions; at the same time, an offset is added to the flight altitude and speed, sample data of the same trajectory at other spatial positions is acquired, and the acquired data is added to the flight parameter data to expand the data volume, thereby achieving the purpose of improving the scope of application of the generated convolutional neural network model.
[0062] Reference Figure 3 , Figure 4 and Figure 5 , Figure 3A schematic diagram of the structure of a convolutional neural network model established for the multi-feature flight maneuver recognition method provided in an embodiment of the present application; Figure 4 A schematic diagram of the structure of a down-sampling ResConv network of a convolutional neural network model established for the multi-feature flight maneuver recognition method provided in an embodiment of the present application; Figure 5 A schematic diagram of the structure of the convolution output channel of the convolutional neural network model established for the multi-feature flight maneuver recognition method provided in an embodiment of the present application.
[0063] In some embodiments, when executing the construction of a convolutional neural network model, the method provided in the present application includes: clustering according to the duration corresponding to all maneuvers in the flight parameter data, and obtaining the widths of multiple corresponding feature map windows of different dimensions; constructing a convolutional neural network model by combining one-dimensional convolution with a residual network, and the convolutional neural network model corresponds to the common parts of multiple feature maps of different dimensions.
[0064] When constructing multiple convolution output channels corresponding to different dimensions after a convolutional neural network model, the method provided in the present application includes: obtaining the size of each window width, and obtaining the dimension of the convolution output channel to be constructed according to the size of each window; and constructing the convolution output channel after the convolutional neural network model according to the dimension of the convolution output channel to be constructed.
[0065] Exemplarily, cluster the duration of all maneuvers in the overall flight parameter data, obtain 9 duration center values, and divide the overall flight parameter data into three groups, and construct the widths of three different dimensional feature map windows. Design the input dimension and output dimension of the neural network based on the provided action type; in the embodiment of the present application, set the input dimension to 3008*25, where 3008 represents the frame time, and 25 represents the flight parameter data corresponding to each frame; set the output dimension to 57*n, where n represents the output parameters of different feature dimensions, and 57 can be decomposed into 3*(16+3), the first 3 represents the three different window widths corresponding to the time window, 16 represents the corresponding action category, and the second 3 represents additional information, namely whether there is action in the window, the center position of the window, and the width of the window.
[0066] By combining one-dimensional convolution with residual network, and based on the common parts extracted from three different dimensional feature maps, a convolutional neural network model is constructed. The input parameter sample of the convolutional neural network model is [number of samples, frame time, flight parameter data], and the dimension is batch_size*3008*25, where batch_size represents the number of samples corresponding to the recognition.
[0067] The convolutional neural network model provided in this application uses convolution to increase the number of channels to 32, and performs batch normalization activation function processing; then, through five times of ResConv downsampling, a convolution kernel of size 3 is used, combined with a residual network, to achieve the purpose of reducing the frame time dimension and increasing the number of channels. The outputs of the last three layers of the constructed convolutional neural network model will be used for branch networks of different dimensional feature extraction.
[0068] In some embodiments, the convolution output channel constructed by the method provided in the present application includes: at least seven layers of neural network convolution layers, wherein the first five layers are used to extract the features of the flight parameter data of the training data set, and the last two layers are used to correspond to the prediction outputs of different dimensions in the prediction feature map.
[0069] The output dimensions corresponding to the convolution output channels are constructed to correspond to the widths of feature map windows of different dimensions. In the embodiment of the present application, there are three convolution output channels. The first channel outputs features of small time dimensions. The output of the last layer of the convolutional neural network model, batch_size*94*1024, is used as input. After seven convolution layers, normalization and activation functions, a large window scale feature map of batch_size*57*94 is finally obtained. The output of the fifth layer of the network is recorded as additional input information for the second channel.
[0070] The second channel uses the extra information in the first channel, and after convolution and upsampling, it is merged with the output of the penultimate layer of the convolutional neural network model as input. It also goes through seven convolutional layers, normalization and activation functions to finally obtain a medium-window scale feature map of batch_size*57*188, and records the output of the fifth layer of the network as additional input information for the third channel.
[0071] The third channel uses the extra information in the second channel, and after convolution and upsampling, it is merged with the third-to-last layer of the main part of the neural network as input. After seven layers of convolution, normalization and activation functions, a small window scale feature map of batch_size*57*376 is finally obtained.
[0072] In some embodiments, when performing decoding and extracting feature data in the corresponding predicted feature map, the feature data in the corresponding predicted feature map provided by the present application includes: prior box center point adjustment parameters, prior box width adjustment parameters, action existence confidence and action type confidence.
[0073] In some embodiments, when performing decoding and converting the flight parameter data into an expression consistent with the form of extracted feature data, the method provided by the present application includes: converting the intermediate moment information and duration information in the flight parameter data corresponding to the feature data into the corresponding feature layer, and obtaining the center point and width of the real box represented in the feature layer; calculating the intersection and union ratio of the real box and the prior boxes of each width obtained by clustering, outputting the prior box with the highest overlap, calculating the offset of the corresponding prior box and the center point of the real box and the width offset; encoding the flight parameter data with action confidence and action type confidence to obtain an expression format consistent with the form of extracted feature data.
[0074] Exemplarily, the flight parameter data obtained by preprocessing is loaded and input into the convolutional neural network model, and the feature map outputs of the three channels can be obtained through the convolutional neural network model; and the feature map outputs corresponding to the three channels are converted into the format of (batch_size, 3, feature_layer, 19), where feature_layer is used to indicate the number of feature layers, corresponding to the feature dimensions of 94, 188 and 376 of the three channel outputs; and 3 represents the three corresponding window widths in the feature layer.
[0075] The feature map data of three sizes are decoded to obtain the prior frame center point adjustment parameters, prior frame width adjustment parameters, action confidence and action type confidence. The action confidence is expressed as whether there is a motorized action in the prior frame. The data here corresponds to 19 of the channel output, that is, three additional information and 16 recognition actions. The decoding method is expressed as follows:
[0076] conf=sigmoid(pred[...,0])
[0077] x = sigmoid(pred[...,1])
[0078] w=pred[...,2]
[0079] pred_cls=sigmoid(prediction[...,3])
[0080] Among them, conf represents the confidence of the existence of action; x represents the adjustment parameter of the center point of the prior box; w represents the adjustment parameter of the width of the prior box; cls represents the confidence of the action type; pred represents the predicted feature map data of three sizes.
[0081] After obtaining the predicted prior frame center position adjustment parameters, width adjustment parameters and action type confidence, use the label corresponding to the flight parameter data in the preprocessed training data set as the target. Convert the intermediate moment information and duration information in the corresponding label to the corresponding feature layer, obtain the center point and width of the real frame represented in the feature layer, and when the real frame completely covers the actual action, calculate the intersection and union ratio between the real frame and the prior frames of each width obtained by clustering, and output the prior frame with the highest overlap; at the same time, calculate the difference between the center position and width of the prior frame and the real frame, and encode the confidence of the existence of action and the confidence of the action type in the same format as the predicted feature map. Finally, get a label with the same format as the prediction result.
[0082] gxs=target[batch_index][:,0]*in_w
[0083] gws=target[batch_index][:,1]*in_w
[0084]
[0085] best_ns=argmax(IOU(true_box,prior_box))
[0086] gxs represents the position of the true center point of the action in the feature layer; gws represents the duration of the action in the feature layer; in_w represents the dimension of the corresponding feature layer; gis represents the frame number of the true frame center point in the feature layer; best_ns represents the index of the prior frame with the highest overlap with the true frame; true_box represents the true frame; prior_box represents the prior frame; IOU is the function for calculating the overlap.
[0087] After that, create the all-zero matrix mask, tx, tw, tconf, box_loss_scale_x of shape (batch_size, 3, feature_layer), the all-one matrix noobj_mask, and the all-zero matrix tcsls of shape (batch_size, 3, feature_layer, 16).
[0088] Among them, mask indicates feature points with targets inside; noobj_mask indicates feature points without targets inside; tx indicates the offset between the center of the prior box with the highest overlap and the real box; tw indicates the adjustment value of the width of the prior box with the highest overlap compared to the width of the real box; tconf indicates the confidence that there is a target action in the prior box; tcls indicates the confidence of each predicted action in the prior box; box_loss_scale_x is the weight of the loss of each segment of the trajectory, where a large weight is selected when calculating a short target action sequence.
[0089] According to the width and feature point position of the a priori frame with the highest overlap, the values of the above matrices are modified. For example, for the 10th data in this batch, for a large window with a feature dimension of 94, the width index of the sample frame with the highest action overlap is calculated to be 2, and the center feature point is located at 34, then:
[0090] mask(10,2,34)=1
[0091] noobj_mask(10,2,34)=0
[0092] tx(10,2,34)=gx-34
[0093] tw(10,2,34)=log(gw / (anchors[2]))
[0094] tconf(10,2,34)=1
[0095] tcls(10,2,34,target)=1
[0096] Among them, gx represents the center point position of the actual box, gw represents the width of the actual box, and anchors represents the different widths corresponding to the current feature layer.
[0097] Then decode the information of all predicted boxes at the same time, calculate the intersection and union ratio with the real box, set the threshold to 0.5, and set all noobj_masks whose intersection and union ratio is greater than the threshold to 0.
[0098] In some embodiments, when performing loss calculation and updating the convolutional neural network model, the method provided by the present application includes: calculating the center point adjustment offset loss through cross entropy; calculating the width adjustment offset loss through mean square error; calculating the action existence confidence and action category confidence loss through cross entropy loss; comprehensively calculating the center point offset loss, width adjustment loss and loss to calculate the action existence confidence and action category confidence loss, and updating the parameters of the constructed convolutional neural network model.
[0099] Exemplarily, after obtaining sample labels that are substantially consistent with the predicted output format, loss calculation and network update are performed.
[0100] Based on the sample data corresponding to the mask, the center offset loss is calculated using cross entropy, which is expressed as follows:
[0101]
[0102] BCE(x,tx)=-log(x)*tx-(1-tx)*log(1-x)
[0103] Based on the sample data corresponding to the mask, the width adjustment loss is calculated using the mean square error. The expression is as follows:
[0104]
[0105] MSE(w,tw)=(w-tw) 2
[0106] Based on the use of cross entropy loss to calculate the action existence confidence and action category confidence loss, the expression is as follows:
[0107]
[0108]
[0109] In summary, the total loss of feature dimension can be expressed as:
[0110] loss = w x *loss_x+w w *loss_w+w conf *loss_conf+w cls *loss_cls
[0111] There are three outputs with different feature dimensions in total. Therefore, after calculating the loss and the average of the three types of feature map outputs, the Adam optimizer is used, the learning rate is set to 1e-5, the gradient of the loss is back-propagated, and the parameters of the convolutional neural network model are updated to reduce the total loss.
[0112] In some embodiments, when executing a prediction of flight parameter data in a test data set through an updated convolutional neural network model, the method provided in the present application includes: inputting the flight parameter data corresponding to the test data set into the updated convolutional neural network model to obtain prediction outputs of three different feature dimensions; calculating the intersection-and-union ratio of the obtained prediction output and the flight parameter data corresponding to the training data set, and selecting the one with the largest intersection-and-union ratio as the final prediction output of the model.
[0113] By using the convolutional neural network model obtained through training and the flight parameter data corresponding to the test data set. The flight parameter data corresponding to the test data set is used as the network input to obtain the prediction output of three different feature dimensions. The intersection and union ratio is calculated with the sample label, and the one with the largest intersection and union ratio is selected as the final prediction output of the model. The recognition accuracy of the test flight action can reach more than 89%.
[0114] The embodiment of the present application first obtains flight parameter data, pre-processes the flight parameter data, and then divides the flight parameter data into a training data set and a test data set for training a convolutional neural network model. Then, feature information is extracted from the flight parameter data corresponding to the training data set by using time convolution windows of different step lengths, and the convolutional neural network model is trained using the acquired multi-feature graph, thereby achieving the purpose of improving the efficiency and accuracy of flight action recognition. At the same time, during the training process, the convolutional neural network model is adjusted by performing loss calculation and calculating the loss parameters obtained by calculation, thereby improving the accuracy and generalization of flight parameter data recognition to a certain extent, thereby further achieving the purpose of improving the accuracy and efficiency of flight action recognition.
[0115] The devices or modules described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. For the convenience of description, the above devices are described in various modules according to their functions. When implementing this application, the functions of each module can be implemented in the same or multiple software and / or hardware. Of course, the module that implements a certain function can also be implemented by combining multiple sub-modules or sub-units.
[0116] Some modules in the apparatus described in the present application can be described in the general context of computer executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.
[0117] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit the present application. Although the present application has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present application.
Claims
1. A multi-feature flight maneuver recognition method, It is characterized in that The method comprises: Acquire and preprocess the flight parameter data, and divide it into a training data set and a test data set; Construct a convolutional neural network model, and construct multiple convolution output channels corresponding to different dimensions after the convolutional neural network model; The output channel includes at least seven convolutional layers, wherein the first five layers are used to extract features of the flight parameter data of the training data set, and the last two layers are used to generate prediction feature maps of different dimensions; the output channel is further set to three channels, which are respectively used to extract features of different time scales to form multiple feature maps of different dimensions; Load the flight parameter data of the training data set into the convolutional neural network model, output the prediction feature graphs corresponding to different dimensions, decode and extract the feature data in the corresponding prediction feature graphs; Load the flight parameter data of the training data set, decode and convert the flight parameter data into an expression consistent with the form of the extracted feature data; when the number of training times reaches the preset number, calculate the loss and update the convolutional neural network model; The flight parameter data in the test data set is predicted through the updated convolutional neural network model; The decoding and converting of the flight parameter data into an expression consistent with the extracted feature data form includes: Convert the intermediate time information and duration information in the flight parameter data corresponding to the feature data into the corresponding feature layer, and obtain the center point and width of the real frame represented in the feature layer; Calculate the intersection-over-union ratio between the true frame and the prior frames of each width obtained by clustering, output the prior frame with the highest overlap, and calculate the offset between the center point of the corresponding prior frame and the true frame and the width offset; The flight parameter data with action confidence and action type confidence are encoded to obtain an expression format consistent with the form of extracted feature data.
2. The multi-feature flight maneuver recognition method according to claim 1, It is characterized in that The acquisition and preprocessing of the flight parameter data comprises: Collect flight parameter data of the aircraft, including flight time, flight altitude, aircraft airspeed and angle of attack; Integrate the acquired flight parameter data of the aircraft and perform data smoothing; Extract the time range corresponding to the maneuver in the flight data, and add category labels to the corresponding maneuvers in the flight data.
3. The multi-feature flight maneuver recognition method according to claim 2, It is characterized in that The obtaining and preprocessing of the flight parameter data also includes: When the amount of flight parameter data obtained is small, the flight altitude change rate, track quaternion change rate, aircraft attitude angle, aircraft attitude angle change rate, and additional state quantities of angular velocity rotation under the body axis system are obtained through differential methods according to the flight time, flight altitude, and track quaternion. At the same time, offsets are added to the flight altitude and speed to obtain sample data of the same trajectory at other spatial positions, and the obtained data is added to the flight parameter data to expand the data volume.
4. The multi-feature flight maneuver recognition method according to claim 1, It is characterized in that The construction of the convolutional neural network model includes: Clustering is performed according to the duration corresponding to all maneuvers in the flight parameter data, and the widths of multiple feature map windows corresponding to different dimensions are obtained respectively; A convolutional neural network model is constructed by combining one-dimensional convolution with a residual network, and the convolutional neural network model corresponds to the common parts of feature maps of multiple different dimensions.
5. The multi-feature flight maneuver recognition method according to claim 4, It is characterized in that The method of constructing a plurality of convolution output channels corresponding to different dimensions after the convolutional neural network model includes: The size of each window width is obtained, and the dimension of the convolution output channel to be constructed is obtained according to the size of each window; the convolution output channel is constructed after the convolution neural network model according to the dimension of the convolution output channel to be constructed.
6. The multi-feature flight maneuver recognition method according to claim 5, It is characterized in that The constructed convolution output channel includes: At least seven convolutional layers of the neural network, of which the first five layers are used to extract the features of the flight parameter data of the training data set, and the last two layers are used to correspond to the prediction outputs of different dimensions in the prediction feature map.
7. The multi-feature flight maneuver recognition method according to claim 1, It is characterized in that In the decoding and extracting of feature data in the corresponding predicted feature map, the feature data includes: a priori box center point adjustment parameter, a priori box width adjustment parameter, action existence confidence and action type confidence.
8. The multi-feature flight maneuver recognition method according to claim 1, It is characterized in that The loss calculation and updating of the convolutional neural network model include: The offset loss is adjusted by calculating the center point through cross entropy; The width adjustment offset loss is calculated by the mean square error; The confidence loss of action existence and action category confidence is calculated through cross entropy loss; The center point offset loss, width adjustment loss, and loss calculation action existence confidence and action category confidence loss are combined to update the parameters of the constructed convolutional neural network model.
9. The multi-feature flight maneuver recognition method according to claim 1, It is characterized in that The predicting of the flight parameter data in the test data set by using the updated convolutional neural network model includes: The flight parameter data corresponding to the test data set is input into the updated convolutional neural network model to obtain prediction outputs of three different feature dimensions; The obtained prediction output is compared with the flight parameter data corresponding to the training data set to calculate the intersection and union ratio, and the one with the largest intersection and union ratio is selected as the final prediction output of the model.
