A method for processing trajectory data for intent recognition

Abstract

The present invention discloses a method for processing trajectory data for intent recognition, belonging to the technical field of trajectory data processing. The present invention includes: collecting trajectory data of an aircraft, performing data preprocessing on the collected trajectory data, and storing the trajectory data in a specified data format; extracting a certain number of trajectory data and performing feature extraction processing; using the dynamic time warping algorithm to discriminate the trajectory similarity, obtaining the trajectory clusters corresponding to the trajectories through hierarchical clustering and adding them as class information to the trajectory data as a piece of training trajectory data; for each trajectory cluster obtained by clustering, calculating the central trajectory of each trajectory cluster through the mutual similarity of the trajectories within the cluster; constructing and training an intent recognition model to obtain the intent recognition result of the trajectory data to be recognized. The present invention can effectively improve the recognition performance of the flight intent of the aircraft.

Description

Technical Field

[0001] The present invention belongs to the technical field of trajectory data processing, and particularly relates to a method for processing trajectory data for intention recognition. Background Art

[0002] As a specific method for predicting the intention of an aircraft based on information such as the flight state, situation, and formation of the aircraft, intention recognition technology is widely used in the aerospace field. Existing intention recognition technologies are mostly based on the multi-dimensional situation of the aircraft. However, in actual situations, trajectory data of the aircraft can usually be obtained, but little information about other aspects can be obtained. The existing intention recognition for the processing and mining of trajectory data is not comprehensive.

[0003] Long short-term memory artificial neural network (LSTM) is used in the field of spatio-temporal data prediction to predict data for a period of time in the future based on historical data. Existing predictions of spatio-temporal data mainly use LSTM and variants of LSTM for prediction. However, LSTM cannot perform parallel operations, which limits the real-time performance of the model, and there are gradient problems for long sequences. Transformer is a model based on the attention mechanism and feed-forward neural network proposed in recent years. Currently, it is widely used in the field of natural language processing (NLP) such as machine translation, question answering systems, text summarization, speech recognition, etc.

[0004] Dynamic Bayesian network is widely used as a traditional intention recognition algorithm. Although it has strict mathematical probability reasoning, the probability matrix is relatively fixed and single, and its self-learning ability is insufficient. Other target intention recognition algorithms rely more on expert experience and knowledge and lack flexibility. One-dimensional convolutional neural network has the advantage of being able to well capture local features of data and identify simple patterns in the data. Then, by utilizing these simple patterns, more complex patterns can be generated in higher-level network layers. Currently, the effect of using one-dimensional convolutional neural network in the processing of sensor-generated sensing data, signal data with a length period, and language data in natural language processing is relatively good. Compared with convolutional neural network, the attention mechanism of Transformer pays more attention to the front-back connection of time series data and performs poorly in mining local features. Summary of the Invention

[0005] The present invention provides a method for processing trajectory data for intention recognition, which is used to improve the recognition performance of the flight intention of an aircraft.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A method for processing trajectory data for intention recognition, the method comprising the following steps:

[0008] Step 1: Collect the trajectory data of the aircraft, preprocess the collected trajectory data, and store the trajectory data in a specified data format. Among them, the trajectory data includes, but is not limited to, the longitude, latitude, altitude of the aircraft, and the data sampling time.

[0009] Step 2: Extract a certain number of trajectory data from the stored trajectory data and perform feature extraction processing.

[0010] Among them, the feature extraction processing includes: generating a feature vector matrix from the extracted trajectory data, calculating the average value of each column of features and the covariance matrix of the features, calculating the eigenvalues and eigenvectors of the covariance matrix, sorting the calculated eigenvalues from small to large and extracting the top K eigenvalues, and then mapping the longitude and latitude in the data to a specified coordinate system. Among them, K > 1.

[0011] Step 3: Set rules for the clustering of trajectory clusters, including the trajectory cluster merging threshold D max and the trajectory cluster number threshold M min , use the dynamic time warping algorithm to judge the trajectory similarity, obtain the trajectory clusters corresponding to the trajectories through hierarchical clustering, and add them to the trajectory data as category information to form a training trajectory data. For each trajectory cluster obtained by clustering, calculate the central trajectory of each trajectory cluster through the mutual similarity of the trajectories within the cluster.

[0012] Step 4: Build and train an intent recognition model.

[0013] The input data of the intent recognition model is the training trajectory data. The input data is subjected to feature information extraction through a one-dimensional convolutional layer, and then connected to a Transformer model through a fully connected layer for intent prediction.

[0014] Input the input data into the intent recognition model, optimize the network parameters of the intent recognition model based on the configured loss function, and stop when the preset training end condition is met. The training end condition can be reaching the maximum number of training times or the prediction accuracy meeting the specified conditions. And use the trained intent recognition model as an intent recognizer.

[0015] Step 5: For the trajectory data to be recognized, calculate its similarity with the central trajectory of each trajectory cluster, and determine the trajectory cluster label of the trajectory data to be recognized based on the trajectory cluster with the highest similarity.

[0016] Input the trajectory data to be recognized and its trajectory cluster label into the intent recognizer, and obtain the intent recognition result of the trajectory data S t to be recognized based on its output.

[0017] The technical solution provided by the present invention at least brings the following beneficial effects:

[0018] The present invention adds information on the dimension of the trajectory cluster to the collected trajectory data, enabling targeted prediction of the trajectory. In the intention recognition process, by combining a one-dimensional convolutional neural network with a Transformer model, the present invention can not only mine the front-back connection of the trajectory data but also fully exploit the local features of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following-described drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0020] Figure 1 is a schematic diagram of the processing process of a trajectory data processing method for intention recognition provided by an embodiment of the present invention;

[0021] Figure 2 is a specific implementation flowchart of a trajectory data processing method for intention recognition provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will further describe the embodiments of the present invention in detail with reference to the drawings.

[0023] According to the actual application scenario of intention recognition based on trajectory data, the present invention specifically sets up a trajectory data processing method for intention recognition. This method can assign trajectories to different trajectory groups or trajectory clusters according to similarity and use the trajectory cluster information as a prior knowledge for subsequent trajectory-based intention recognition (intention prediction). Through the processed trajectory data, intention recognition can be carried out quickly and with high accuracy.

[0024] As Figure 1 and Figure 2 shown, a trajectory data processing method for intention recognition provided by an embodiment of the present invention includes the following steps:

[0025] Data acquisition: Collect the data sent down by the aircraft detection device.

[0026] Data storage and processing: The data collected is raw sensor data. Therefore, data storage and processing is responsible for storing the raw sensor data collected after the first data preprocessing (such as unifying data units, data formats, detecting outliers, normalizing, etc.) into the memory. At the same time, it is responsible for extracting specific data from the database for corresponding processing.

[0027] Data Feature Extraction: Generate a feature vector matrix from the data extracted from the data storage, calculate the average value of each column feature and the covariance matrix of the features. Calculate the eigenvalues and eigenvectors for the covariance matrix, and sort the calculated eigenvalues from smallest to largest and extract the top k eigenvalues. At the same time, map the longitude and latitude in the data to a specified coordinate system.

[0028] Trajectory Clustering: The clustering uses the Dynamic Time Warping (DTW) algorithm to discriminate the trajectory similarity, and set the trajectory cluster merging threshold D max and the trajectory cluster number threshold M min Set rules for the clustering of trajectory clusters. Obtain the trajectory clusters corresponding to the trajectories through hierarchical clustering and add them as category information to the trajectory data. For the obtained trajectory clusters, calculate the central trajectory through the mutual similarity of the trajectories within the cluster.

[0029] Intention Prediction: Intention prediction includes the processing of training data and the intention prediction of the aircraft. The processing of training data is based on the trajectory data after the trajectory clustering processing step, that is, taking the trajectory cluster label to which the current trajectory belongs as an attribute information to obtain the trajectory data with trajectory clusters. Then train with the data in the expert knowledge base and label the training data. Send the labeled data into the 1DCNN+Transformer model for training. First, the data passes through the 1DCNN convolutional layer (i.e., a one-dimensional convolutional layer) to extract information from the input features, and then passes through a fully connected layer to be connected to the Transformer model for intention prediction.

[0030] Among them, the attention feature interaction part of the Transformer model adopts the multi-head self-attention mechanism, and multiple multi-head self-attention layers can be stacked according to the situation. Use h different linear transformations to project the query vector Q, the key vector K, and the value vector V, splice the results of different attention mechanisms, obtain the output of the multi-head attention mechanism, and after normalization processing, finally obtain the final prediction result through the fully connected feed-forward neural network.

[0031] As a possible implementation manner, in the embodiments of the present invention, the data format of the data collection is shown in Table 1:

[0032] Table 1

[0033]

[0034]

[0035] Among them, (Lon,Lat) represents the longitude and latitude of the aircraft, Alt represents the altitude, and Time represents the time.

[0036] The processing steps of trajectory clustering include:

[0037] Step 1: Perform original data cleaning and transformation operations:

[0038] Step 1.1: Clean the original data of the periodically detected aircraft, and delete the singular points;

[0039] Step 2: Perform data preprocessing operations on the aircraft cluster:

[0040] Step 2.1: Extract the key features of the aircraft data, that is, generate a feature vector matrix from the preprocessed trajectory data, calculate the average value of each column of features and the covariance matrix of the features. Calculate the eigenvalues and eigenvectors for the covariance matrix, and sort the calculated eigenvalues from smallest to largest and extract the top k eigenvalues.;

[0041] Step 2.2: Map the longitude and latitude values of each aircraft to the coordinate system. Let the collected longitude and latitude of the aircraft be (Lon, Lat), set the coordinate of the origin of the coordinate system as (Lon_o, Lat_o) with the value of (31.0363836, 103.6507575), and the mapped coordinate of the coordinate system is set as (x, y). The mapping rule is as follows:

[0042] x = (Lon - Lon_o) * 3600 * 30.9;

[0043] Step 3: Perform aircraft trajectory clustering:

[0044] Step 3.1: Use the DTW algorithm to calculate the similarity between trajectories, and obtain the similarity distance matrix between trajectories. L(S i , S j ) represents the similarity value between the S i trajectory and the S j trajectory. The smaller this value, the more similar the two trajectories are;

[0045] Step 3.2: Let the two trajectory clusters be C n , C m , where C n contains s1 trajectories, C m contains s2 trajectories, and the similarity of the trajectory clusters of C n and C m is calculated as follows: The calculation formula is as follows:

[0046]

[0047] The smaller this value, the closer the two trajectory clusters are;

[0048] Step 3.3: Set the trajectory cluster merging threshold D max , when indicates that the two trajectory clusters do not meet the merging requirements;

[0049] Step 3.4: Set the trajectory cluster number threshold M min , when the number of trajectories in the cluster is less than M min , it is not classified as a trajectory cluster in the final clustering result;

[0050] Step 3.5: Perform hierarchical clustering;

[0051] Step 3.6: Evaluate the clustering result and perform optimization;

[0052] Step 3.7: Through the obtained trajectory cluster information.

[0053] In this embodiment, the trajectory cluster data and trajectory data formats are shown in the following table:

[0054] Table 2

[0055]

[0056] Table 3

[0057]

[0058] Step 3.8: For each trajectory cluster C i , obtain the clustering center of each trajectory cluster through the formula;

[0059]

[0060] where n represents the trajectory number, and C i {n) represents the trajectory numbered n in the i-th trajectory cluster.

[0061] That is, the clustering center track O of the i-th trajectory cluster i is the track with the smallest sum of similarities with other tracks in the cluster;

[0062] Step 4: Perform aircraft intention prediction:

[0063] Step 4.1: Based on the obtained trajectory data with trajectory clusters as training data, label the training data through the expert database, and the configured data format is shown in Table 4:

[0064] Table 4

[0065]

[0066] where type represents the trajectory cluster label and intent represents the intention label.

[0067] Step 4.2: Send the training data into the 1DCNN+Transformer model (i.e., the intention recognition model) for training;

[0068] Step 4.3: Evaluate the prediction results (flight intentions) output by the 1DCNN+Transformer model through the expert database;

[0069] Step 4.4: Optimize the model parameters and perform iterative training, and stop when the preset training end condition is met. The training end condition can be reaching the maximum number of training times or the prediction accuracy meeting the specified conditions;

[0070] Step 4.5: Use the trained intention recognition model as an intention recognizer and store it;

[0071] Step 5, for the trajectory data S to be recognized t , first calculate its similarity with the cluster center track O of each trajectory cluster i , obtain the trajectory cluster with the highest similarity to it, and obtain the trajectory data S corresponding to the trajectory cluster t trajectory cluster label;

[0072] Input the trajectory data S t and its trajectory cluster label into the intention recognizer, and obtain the intention recognition result of the trajectory data S to be recognized based on its output t .

[0073] 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 foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0074] The above are only some embodiments of the present invention. For those of ordinary skill in the art, without departing from the inventive concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A method for processing trajectory data for intent recognition, characterized in that, It includes the following steps: Step 1: Collect the trajectory data of the aircraft, perform data preprocessing on the collected trajectory data, and store the trajectory data in a specified data format. Among them, the trajectory data includes the longitude, latitude, altitude, and data sampling time of the aircraft; Step 2: Extract a certain number of trajectory data from the stored trajectory data and perform feature extraction processing; Among them, the feature extraction processing includes: generating a feature vector matrix from the extracted trajectory data, calculating the average value of each column of features and the covariance matrix of the features, calculating the eigenvalues and eigenvectors of the covariance matrix, sorting the calculated eigenvalues from smallest to largest and extracting the top K eigenvalues, and then mapping the longitude and latitude in the data to a specified coordinate system. Among them, K>1; Step 3: Use the dynamic time warping algorithm to judge the trajectory similarity, obtain the trajectory clusters corresponding to the trajectories through hierarchical clustering, and add them as class information to the trajectory data as a training trajectory data. For each trajectory cluster obtained by clustering, calculate the central trajectory of each trajectory cluster through the mutual similarity of the trajectories within the cluster; Step 4: Construct and train an intention recognition model; The input data of the intention recognition model is the training trajectory data. The input data undergoes feature information extraction through a one-dimensional convolutional layer, and then is connected to a Transformer model through a fully connected layer for intention prediction; Input the input data into the intention recognition model, optimize the network parameters of the intention recognition model based on the configured loss function, and stop when the preset training end condition is met. The training end condition can be reaching the maximum number of training times or the prediction accuracy meeting the specified conditions. And use the trained intention recognition model as an intention recognizer; Step 5: For the trajectory data to be recognized, calculate its similarity with the central trajectory of each trajectory cluster, and determine the trajectory cluster label of the trajectory data to be recognized based on the trajectory cluster with the highest similarity; Input the trajectory data to be recognized and its trajectory cluster label into the intention recognizer, and obtain the intention recognition result of the trajectory data S to be recognized based on its output. t ​ 2. The method according to claim 1, characterized in that, In Step 2, mapping the longitude and latitude in the data to a specified coordinate system specifically means: Define the longitude and latitude of the collected aircraft as (Lon, Lat), define the coordinate origin of the longitude and latitude in the collected data as (Lon_o, Lat_o), and the longitude and latitude of the aircraft after mapping as (x, y); The mapping relationship is: x = (Lon - Lon_o) * 3600 * 30.9; 3. The method according to claim 1, characterized in that, Step 3 further includes performing a merging process on the trajectory clusters obtained through the clustering process: According to the formula calculate the similarity between the trajectory cluster as C n and C m When merge the trajectory clusters C n and C m where D max represents the trajectory cluster merging threshold.

4. The method according to claim 1, wherein Step 3 further includes: for each trajectory cluster obtained currently, when the number of trajectories within the cluster is less than a preset trajectory cluster number threshold M min discard the current trajectory cluster.

5. The method according to claim 1, wherein In Step 3, the central trajectory of each trajectory cluster is: Among them, O i represents the central trajectory of the i-th trajectory cluster, n represents the trajectory number, and C i (n), C i (m) represent the trajectories numbered n and m in the i-th trajectory cluster respectively, s represents the maximum trajectory number within the i-th trajectory cluster, and L() represents the similarity between trajectories.

Images