Contrast learning sample construction and prediction method and system for flight arrival time prediction

Abstract

The invention belongs to the field of big data, and particularly relates to a comparative learning sample construction and prediction method and system for flight arrival time prediction. The method comprises the following steps: acquiring trajectory data and meteorological observation data of each flight of an airport and preprocessing the trajectory data and the meteorological observation data; and constructing a time sequence sample by using the preprocessed time axis aligned trajectory data and flight feature data of the meteorological observation data. Performing sample phase division according to track features, meteorological features and residual arrival time in the end timestamps of the time sequence samples, and enabling the time sequence samples in the same phase to be weak positive samples. Calculating a feature combination distance between the time sequence samples in the same phase; and screening a specified number of nearest neighbors as strong positive samples of the nearest neighbors. And using the data set containing the positive sample label to carry out mixed contrast learning training on the flight prediction model. According to the method, the problems of insufficient prediction precision and generalization of the existing flight arrival time prediction are solved.

Description

Technical Field

[0001] This invention belongs to the field of big data, specifically relating to a comparative learning sample construction method for predicting flight arrival times, and a corresponding method and system for predicting flight arrival times. Background Technology

[0002] With the continuous expansion of global civil aviation transport and sustained growth in flight demand, the aviation operation system is characterized by high flight density, resource constraints, and frequent operational disruptions. Against this backdrop, more accurate flight arrival time prediction has become a crucial foundation for improving airspace and airport operational efficiency. On the one hand, air traffic control needs to create schedules based on the estimated arrival times of aircraft at key constraint points to reduce guidance and waiting / circling time and improve airspace capacity utilization. On the other hand, airport collaborative decision-making mechanisms emphasize improving operational efficiency, resilience, and predictability of flight operations through more accurate, timely, and consistent information sharing. Therefore, under the conditions of increasing flight volume and operational uncertainty, researching and implementing high-precision, generalizable arrival time prediction technology is of great significance for reducing delays, optimizing resource allocation, and improving overall operational efficiency.

[0003] In recent years, numerous studies on flight arrival time prediction have employed machine learning or deep learning methods, utilizing multi-source information such as ADS-B flight data, flight schedules, and weather data to predict remaining arrival times. However, most studies focus on single-airport datasets. Each airport operates different routes and carries different airlines, resulting in significant differences in flight operation characteristics across different airlines and aircraft types. Single-airport datasets limit the generalization ability of models, leading to insufficient transferability. To reduce the dependence of some supervised learning algorithms on high-quality labels and improve model transferability, contrastive learning has begun to be introduced when processing time series data such as flight data.

[0004] Contrastive learning learns data representation methods by comparing positive and negative sample pairs. Its core idea is to use a loss function to group similar instances closer together and dissimilar instances further apart in the learned embedding space. However, flight data samples are highly sensitive to how positive samples are constructed. Conventional contrastive learning methods that use physical priors such as pruning, jittering, and permutation to construct positive samples are not always effective for time-series data like flight data, which involves phase transitions (cruise, approach, etc.), abrupt changes, and anomalies. This can easily introduce spurious positive samples, weakening the quality of the model's learned representation and the stability of downstream prediction tasks. Therefore, there is an urgent need for a technical solution that can combine the physical / operational patterns and environmental factors of flight trajectories to construct more reliable positive-negative sample relationships and learn physically consistent temporal representations, thereby improving the accuracy of flight arrival time prediction and cross-scenario generalization capabilities. Summary of the Invention

[0005] To address the limitations of existing neural network-based flight arrival time prediction methods, which suffer from insufficient prediction accuracy and generalization ability due to the quality of the dataset, this invention provides a comparative learning sample construction method for flight arrival time prediction, along with a corresponding flight arrival time prediction method and system.

[0006] This invention is achieved using the following technical solution: A method for constructing contrastive learning samples for flight arrival time prediction, comprising: Acquire historical arrival flight trajectory data and meteorological observation data for a designated airport; preprocess the acquired data to construct the feature data required for flight arrival time prediction.

[0007] The trajectory data is categorized by flight number and sorted by time, retaining the tail segment of the trajectory data with remaining arrival times not exceeding a specified value for modeling. A time series sample of a preset length is constructed using a sliding window method, and data in the window sample that is shorter than the preset length is pre-padded; the time series sample is then standardized by ignoring the padded values.

[0008] Based on the trajectory characteristics, meteorological characteristics, and remaining arrival times in the final timestamps of any time series sample, the sample phase is divided, and time series samples within the same phase are considered as weak positive samples to each other. The calculation of any two time series samples within the same phase... i and j Feature combination distance between D ( i , j ); Select multiple nearest neighbors of each time series sample as its strong positive samples, and then form a global strong positive sample index matrix.

[0009] The present invention also includes a method for predicting flight arrival times, comprising: A two-layer LSTM and a one-layer fully connected layer are used as the contrastive learning encoder; the contrastive learning encoder is used to learn from the input time sample sequence. X i Generate intermediate vector representation h i The dataset containing a large number of weak positive samples and strong positive samples with labels, obtained by using the contrastive learning sample construction method for flight arrival time prediction as described above, is used to train the aforementioned contrastive learning encoder using hybrid contrastive learning.

[0010] In the hybrid contrastive learning training phase, the padding values ​​in the input time series samples are first masked, and then... h i A subsequent projection head and L2 normalization are performed to obtain embedding vectors specifically for contrastive learning.z i The hybrid contrastive training phase employs a loss function that simultaneously incorporates physical feature phase contrastive learning loss and feature combination distance contrastive learning loss. .

[0011] After the hybrid contrastive learning is completed, a fully connected layer is added after the contrastive learning encoder as a regression output layer to build the prediction model. The samples in the collected dataset are sorted by date, and all samples from the most recent 20% of dates are used as the validation set, with the remainder used as the training set; the MAE loss function is then applied. Train the flight arrival time prediction model.

[0012] The trained prediction model is used to predict flight arrival times based on the input time feature sequence.

[0013] As a further improvement of this invention, the hybrid contrastive learning training includes phase contrastive learning based on physical features and distance contrastive learning based on feature combination. In the former, each time series sample in the dataset is considered a positive sample if its weak positive samples are positive, and the rest are negative samples; in the latter, its strong positive samples are positive samples, and the rest are negative samples.

[0014] The present invention also includes a flight arrival time prediction system, which includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it creates a prediction model trained in the aforementioned flight arrival time prediction method and outputs the predicted flight arrival time based on the input time series samples.

[0015] The technical solution provided by this invention has the following beneficial effects: This invention fully considers the characteristics of flight data itself and designs a more reasonable method for constructing positive and negative sample pairs for comparative learning tasks used in flight arrival time prediction. This method divides the sliding window samples of the flight's tail-segment trajectory time series data into phases based on flight pressure altitude, ground speed, distance, weather conditions, and remaining arrival time. Each sample is structurally more similar to samples in the same phase and less similar to samples in other phases; therefore, this invention considers samples in the same phase as weak positive samples. Furthermore, considering that each phase division is still relatively coarse, blindly treating all samples in the same phase as equally weighted positive samples is inappropriate. Therefore, this invention also calculates the differences in partial trajectory features and weather features between samples using weighted calculations and calculates the feature combination distance (PC-MTD) between different samples, thereby uncovering more similar strong positive samples.

[0016] This invention builds upon traditional deep learning for flight arrival time prediction by combining contrastive learning and the LSTM deep learning algorithm. First, contrastive learning is used to train an encoder (deep learning model) capable of learning more general and stable time-series representation vectors. This encoder is then used to encode the input samples, followed by a regression prediction task. Under the same sample quality conditions, this results in more stable predictions and a more generalizable prediction model. It addresses the problem of poor generalization in existing prediction models, which require separate retraining for different flight data samples.

[0017] Based on the new data and model architecture, this invention also designs a two-stage training strategy. The first stage, hybrid contrastive learning, introduces similarity constraints between aircraft trajectory morphology and wind field conditions. This mechanism more accurately encodes key factors affecting the remaining arrival time into the embedding space, thereby improving the ability of the vector representation learned by the contrastive learning encoder to distinguish complex operating conditions. The second stage uses the MAE loss function to train the overall prediction model including the contrastive learning encoder. Specifically, the distinction between weak and strong positive samples is achieved by adjusting the weights of the loss function.

[0018] This invention addresses the problem that general time series contrastive learning relies on data augmentation methods such as pruning, jittering, and permutation to construct positive samples in flight arrival time prediction scenarios. If the data augmentation is not handled properly, it can easily create pseudo-positive samples that do not conform to flight dynamics and air traffic control constraints. This invention also significantly improves the prediction accuracy and generalization of flight arrival times. Attached Figure Description

[0019] Figure 1 This is a flowchart of the steps in the comparative learning sample construction method for flight arrival time prediction provided in Embodiment 1 of the present invention.

[0020] Figure 2 This is a schematic diagram of the principle of the positive sample pair partitioning method.

[0021] Figure 3 This is a flowchart of the flight arrival time prediction method provided in Embodiment 2 of the present invention.

[0022] Figure 4 This is a schematic diagram illustrating the two-stage training principle of the contrastive learning encoder and its prediction model provided in Embodiment 2 of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] Example 1

[0025] To address the issue that flight arrival time prediction methods based on contrastive learning are highly sensitive to the quality of positive and negative sample pairs used in the training phase, and that existing technologies lack the ability to extract a sufficient number of high-quality positive samples from massive amounts of historical data related to flight operations, this embodiment provides a method for constructing contrastive learning samples for flight arrival time prediction. For example... Figure 1 and Figure 2 As shown, the method includes the following steps: S1: Obtain historical arrival flight trajectory data and meteorological observation data for the designated airport.

[0026] The trajectory data of each historical arriving flight at each airport can be automatically collected by the ADS-B (Automatic Dependent Surveillance-Broadcast) system. The ADS-B system can periodically and automatically collect and broadcast information such as the real-time position, altitude, speed, heading, and identification number of aircraft, achieving air traffic status monitoring without manual operation. In this embodiment, the trajectory data of each flight collected by the ADS-B system includes: timestamp, 24-bit ICAO address, latitude, longitude, ground speed, heading angle, vertical speed, flight call sign, whether on the ground, alarm status, special location marker, transponder code, barometric altitude, geometrical altitude, ADS-B transmitter serial number, and the hour period to which the data was collected, etc.

[0027] Meteorological observation data for each airport is obtained through published civil aviation meteorological reports, which are divided into two categories: meteorological observation reports and meteorological forecast reports. From a categorized perspective, civil aviation meteorological reports can be further divided into routine observation reports, special observation reports, and weather forecasts. In this embodiment, the meteorological observation data collected for any airport generally includes: observation time, airport ICAO code, wind direction, wind speed, visibility, cloud cover, cloud height, air temperature / dew point temperature, surface air pressure, etc.

[0028] S2: Preprocess the acquired data to construct the feature data required for the flight arrival time prediction task.

[0029] In this embodiment, all the collected interrelated trajectory data and meteorological observation data need to be cleaned and then processed according to the specified preprocessing method to convert the raw data into feature data with a special format required for the flight arrival time prediction task.

[0030] Specifically, in practical applications, the preprocessing of trajectory data and meteorological observation data generally includes the following steps: (1) Convert the timestamps of the trajectory data to Coordinated Universal Time, delete the entire row of data with empty timestamps, and divide it into two columns according to timestamp and flight number to perform deduplication.

[0031] (2) Delete all abnormal data in the data such as ground speed and air pressure altitude. In this embodiment, abnormal data includes data points that significantly exceed reasonable thresholds or have abrupt changes in the time series data.

[0032] (3) The spherical distance from the flight trajectory point to the target airport is calculated using the semi-sine formula based on the coordinates of the flight trajectory point. After the calculation is completed, the distance corresponding to each timestamp of each flight is added to the feature.

[0033] Specifically, let the target airport coordinates be (x0, y0), and the coordinates of the trajectory point of flight z at timestamp t be... Then, the formula for calculating the spherical distance from flight z to the target airport is as follows: ; Where R is the Earth's radius.

[0034] Since the heading angle is a periodic variable, directly inputting the angle value will cause discontinuous jumps at the boundary between 0° and 360° and introduce spurious differences. Therefore, this invention uses the heading angle... Mapped to and Two-dimensional features are used to achieve continuous representation of periodic variables and improve the stability and generalization ability of sequence modeling.

[0035] (4) Perform second-by-second resampling and conditional interpolation on each flight's data.

[0036] In this embodiment, the time range of the time series trajectory data of any flight z is [ ], then the second-by-second timeline Defined as:

[0037] Define the gap ratio for:

[0038] in, This represents the number of timestamps on the original trajectory of the flight.

[0039] set up As the threshold for the gap ratio, when At that time, no second-by-second resampling or interpolation will be performed, and all data for this flight will be deleted to avoid "over-fabrication" under conditions of large gaps; when When there are missing segments of no more than 10 seconds in the time series trajectory data of the flight, linear interpolation is performed on the features of the data that are continuous numerical fields (such as barometric altitude, ground speed, longitude, latitude, etc.).

[0040] (5) Perform a 3-second window median filter on the continuous numerical fields within each flight to suppress spikes and jitter noise. In this embodiment, the continuous numerical fields include the fields corresponding to the altitude column, speed column, and heading angle column.

[0041] (6) Resample the meteorological observation data minute by minute and then align it with the trajectory data second by second. In this embodiment, the data density of meteorological observation data is usually at the minute level, while the trajectory data is at the second level. After resampling the meteorological observation data minute by minute in this embodiment, interpolated second-level meteorological observation data can be obtained. Then, the two types of data can be aligned with the time axis according to the timestamp, so that the trajectory data for each second has corresponding meteorological observation data. The two together constitute the second-level feature data for predicting flight arrival time.

[0042] S3: Construct the required time series samples using preprocessed flight feature data that includes time-axis aligned trajectory data and meteorological observation data.

[0043] In this embodiment, the trajectory data is first classified by flight number and sorted by time, retaining the tail segment of trajectory data with remaining arrival times not exceeding a specified value for modeling. Then, a time series sample of a preset length is constructed using a sliding window method; data in window samples shorter than the preset length are pre-padded; and the time series samples are standardized by ignoring padding values. Similarly, padding values ​​and non-finite values ​​are ignored when calculating the mean and variance for each feature dimension.

[0044] Specifically, the construction process of time series samples is as follows: (1) Calculate the remaining arrival time for each timestamp of the trajectory sequence of flight z. y z ( t ), and only retain those that meet the requirements. y z ( t )≤y max Trajectory data fragments. Where y max This is the preset maximum length of the final segment of a flight.

[0045] (2) Set the sliding step size Δ, and select the end timestamp of the time series sample of each flight in chronological order from the set of trajectory points that meet the segment length in the previous step. Then the end timestamp of the sliding window sample selection for each flight should satisfy: ; in, For the first i The start timestamp of the trajectory data of flight z in the last 30 minutes to which each time series sample belongs. For the first iThe last timestamp of the trajectory data of flight z to which each time series sample belongs; k This represents the number of sliding window iterations during the time series sample generation phase. N It represents the set of natural numbers.

[0046] (3) For any 1st i End timestamps of time series samples The starting point of its sliding window for: ; in, T Indicates the sample length of the preset sliding window; The final extracted window fragment is:

[0047] in, L The effective length of the sliding window sample; F For feature dimensions.

[0048] (4) When the sliding window sample length L < T At that time, the window sample is filled in front ( T - L The padding value for each time step is set to PAD-VAL = -999.

[0049] S4: Divide the sample phase based on the trajectory features, meteorological features and remaining arrival time in the end timestamp of any time series sample, and treat each time series sample in the same phase as a weak positive sample.

[0050] Specifically, the method for determining weak positive samples in this embodiment includes the following: (1) Extract any number of... i End timestamps of time series samples The distance in the corresponding trajectory features and air pressure altitude The data, and the flight phases were determined. phase 1.

[0051] In detail, the flight phase in this embodiment phase 1 is divided into 3 types, with values ​​ranging from 0 to 2. The flight phase corresponding to any time series sample. phase The value of 1 is calculated using the following formula: ; In the above formula, D1 and H1 are the distance threshold to the target airport and the flight pressure altitude threshold corresponding to phase 1, respectively, and D2 and H2 are the distance threshold to the target airport and the flight pressure altitude threshold corresponding to phase 2, respectively.

[0052] (2) Extract any number of... i End timestamps of time series samples Wind direction in the corresponding meteorological characteristics Wind speed Heading angle The data, and the delineation of meteorological phases. phase 2.

[0053] In detail, this embodiment defines the relative angle between wind direction and flight heading angle. for: ; The relative angles are expressed in radians. : ; Then, the wind speed component along the heading direction for: ; Wind speed component in the vertical heading direction for: ; Meteorological phase in this embodiment phase 2 is divided into 4 types, with values ​​ranging from 0 to 3. The meteorological phase corresponding to any time series sample. phase The value of 2 is calculated using the following formula: ; In the above formula, W It is the weak wind speed threshold; and They represent the first i End timestamps of time series samples The wind speed components in the parallel and perpendicular directions of flight; phase The value 2 can be 0, 1, 2, or 3, representing weak wind, tailwind, headwind, and crosswind respectively.

[0054] (3) Extract any number of... i End timestamps of time series samples The remaining arrival time is and define the arrival time phase. phase 3.

[0055] Arrival time phase in this embodiment phase 3 is divided into 6 types, with values ​​ranging from 0 to 5. The remaining arrival time phase corresponding to any time series sample. phase The value of 3 is calculated using the following formula: .

[0056] (4) Calculate the label of each time sample sequence using the following formula. label Based on the labels of each time sample sequence, any two time sample sequences with the same label are considered weak positive samples of each other. .

[0057] In this embodiment, corresponding label information is added to each time series sample according to the partitioning results, and the samples are grouped. Time series samples with the same label are samples within the same group, and they can be regarded as weak positive samples to each other during the comparative learning phase.

[0058] S5: Calculate any two time series samples within the same phase. i and j Feature combination distance between D ( i , j ); Select multiple nearest neighbors of each time series sample as its strong positive samples, and then form a global strong positive sample index matrix.

[0059] In detail, the method for dividing strong positive samples in this embodiment includes the following process: Given the first i Time series samples X i for: ;

[0060] in, These are the feature vectors corresponding to the s-th, s+1-th, ..., e-th timestamps contained herein; The first The feature vector of each timestamp contains the 1st, 2nd, ..., Fth dimension features.

[0061] Define any first i The time steps of each time series sample t The formula for determining whether a time step is valid is as follows: ; in, m i ( t If )=1, it indicates that it is a valid time step. m i ( t If )=0, it indicates that the time step is invalid.

[0062] Corresponding to any two time sample sequences i and j The common effective time step set of both for: ; For any ,definition i and j The f dimensional feature difference is : ; Based on this, set This is the weight vector for the trajectory features. ; This is the weight vector of meteorological features. ; F 1 and F 2 represents the dimensions of trajectory features and meteorological features, respectively.

[0063] Then any two time series samples within the same phase i and j Feature combination distance between D ( i , j The formula for calculating ) is: ; In the above formula, express i and j The common effective time step set among them; express i and j In the t The feature combination distance at each time step; and They are respectively i and j In the t The difference in trajectory features and the difference in meteorological features at each time step; and These represent the weights of trajectory feature difference and meteorological feature difference in the feature combination distance, respectively. In this embodiment, .

[0064] It should be noted that if Then define , representing time series samples i and j PC-MTD values ​​cannot be calculated between these points. D max This is the preset maximum distance.

[0065] in, and The calculation formula is as follows: ; In the above formula, For the trajectory features, the first f dimensional weight vector, The first of the meteorological characteristics f A weight vector of dimension; express i and j In the f Feature difference in dimensional trajectory features; express i and j In the f The difference in meteorological characteristics.

[0066] In practical applications, based on the phase grouping in step S4, this step can calculate the PC-MTD value between samples within each group. The PC-MTD value is used to measure sample similarity; the smaller the value, the more similar the two samples are. g Number of positive samples with strong internal strength K g The definition is as follows: ; in, n g For group g Number of samples within, k max and k min These are the upper and lower limits of the number of strongly positive samples within each group. It is the pre-defined proportion of strongly positive samples in the group.

[0067] According to the calculations within the group D ( i , j Sort by size from smallest to largest, and take the first few. K g Sample , , ..., As time series samples i Strong positive samples, then i Strong positive sample set P s ( i ) is represented as: ; in, .

[0068] For groups with fewer than n g The group does not perform PC-MTD value calculation and does not participate in PC-MTD comparative learning.

[0069] Finally, based on the determination of whether they are strong positive samples, this embodiment can generate a corresponding global strong positive sample index matrix.

[0070] Example 2

[0071] Based on the positive sample data constructed in Example 1, this example further provides a method for predicting flight arrival times, such as... Figure 3 As shown, the method includes the following steps: S01: Construct a contrastive learning encoder by performing hybrid contrastive learning training on a dataset containing a large number of time sample sequences of labeled weak positive samples and strong positive samples from Example 1.

[0072] Specifically, such as Figure 4 As shown, this embodiment uses a two-layer LSTM and a fully connected layer as the contrastive learning encoder. At this stage, a projection head and an L2 normalization layer are added after the contrastive learning encoder. The contrastive learning encoder is used to learn from the input time sample sequence. X i Generate intermediate vector representation h i In the hybrid contrastive learning training phase, the padding values ​​in the input time series samples are first masked, and then passed sequentially through two LSTM layers and a fully connected layer to obtain an intermediate vector representation. ; and in obtaining h i A subsequent projection head and L2 normalization are performed to obtain embedding vectors specifically for contrastive learning. z i .

[0073] The first LSTM layer input is a sequence of time samples. X i The output is H i .

[0074] , in, For sliding window samples timestamp The representation vector, , The output dimension of the vector is represented by each preset timestamp.

[0075] The input to the second LSTM layer is H i The output is the representation vector of the last time step of the input sample. , , The output dimension is the preset representation vector.

[0076] In this embodiment, the hybrid contrastive learning training includes phase contrastive learning based on physical features and distance contrastive learning based on feature combination. In the former, each time series sample in the dataset is considered a positive sample if its weak positive samples are positive, and the rest are negative samples; in the latter, its strong positive samples are positive samples, and the rest are negative samples.

[0077] In the hybrid contrastive training phase, a loss function is used that simultaneously incorporates physical feature phase contrastive learning loss and feature combination distance contrastive learning loss. The expression for the loss function is as follows: ; in, and They are the first i The physical feature phase contrast learning loss and feature combination distance contrast learning loss are used for each time series sample. and They are and Weights in the total loss function; P label ( i ) indicates the first i A set of weak positive samples from a time series sample; P s ( i ) indicates the first i A set of strongly positive samples from a time series sample; z i and z j For the first i and the j Embedded vectors of time sample sequences; This is a scaling factor used to adjust the distribution of similarity between samples; the smaller the value, the more sensitive the loss function is to differences in similarity, and the closer the positive samples are; the larger the value, the smoother the similarity between samples.

[0078] S02: Build a prediction model and train it for regression prediction.

[0079] After the mixed contrastive learning is completed, such as Figure 4 As shown, this embodiment retains the model parameters of the contrastive learning encoder after the corresponding training is completed, and adds a fully connected layer after the contrastive learning encoder as a regression output layer to construct a prediction model.

[0080] The samples in the collected dataset are sorted by date, and all samples from the most recent 20% of dates are used as the validation set, with the remainder used as the training set; the MAE loss function is then applied. Train the flight arrival time prediction model.

[0081] MAE Loss Function for: ; in, It is the first i A time series sample in The actual remaining arrival time; It is the predicted remaining arrival time output by the prediction model. M It is the number of samples in a training batch.

[0082] S03: Use the trained prediction model to predict flight arrival times based on the input time feature sequence.

[0083] Example 3

[0084] Based on the solution in Embodiment 2, this embodiment further provides a flight arrival time prediction system, which includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it creates a prediction model trained in the aforementioned flight arrival time prediction method and outputs the predicted flight arrival time based on the input time series samples.

[0085] The flight arrival time prediction system provided in this embodiment is essentially a computer device. This computer device can be a smartphone, tablet, laptop, desktop computer, rack server, blade server, tower server, or cabinet server (including standalone servers or server clusters composed of multiple servers), etc., capable of executing programs.

[0086] The computer device in this embodiment includes, but is not limited to, a memory and a processor that can be interconnected via a system bus. In this embodiment, the memory (i.e., the readable storage medium) includes flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory can be an internal storage unit of the computer device, such as the hard disk or RAM of the computer device. In other embodiments, the memory can also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Of course, the memory can also include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is typically used to store the operating system and various application software installed on the computer device. Furthermore, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0087] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of a computer device.

[0088] Simulation test

[0089] To verify the technical effectiveness of the comparative learning sample construction method for flight arrival time prediction and the corresponding flight arrival time prediction method provided by this invention, the technicians formulated an experimental plan, constructed positive samples based on the collected relevant data, and used them to train and test the constructed flight prediction model. Finally, the performance of the trained flight prediction model was evaluated.

[0090] I. Construction of Dataset and Positive Samples

[0091] The time-series data related to flight trajectories in this experiment were sourced from all arriving flights at London Gatwick Airport from June 1, 2024 to June 19, 2024. The data includes: timestamp, 24-bit ICAO address, latitude, longitude, ground speed, heading angle, vertical speed, flight call sign, ground status, alarm status, special location markers, transponder code, barometric altitude, geometric altitude, ADS-B transmitter serial number, and the hour period to which the data was collected. Weather data was sourced from Gatwick Airport meteorological observation reports, including observation time, airport ICAO code, wind direction, wind speed, visibility, cloud cover, cloud height, air / dew point temperature, and surface pressure.

[0092] Data cleaning and fusion are performed on the original flight trajectory time series data and meteorological data, and corresponding feature construction is carried out. In this embodiment, data with ground speeds greater than 420 knots and air pressure altitudes greater than 40,000 feet are identified as obviously abnormal data, and these data rows are deleted entirely. The designated airport in this embodiment is London Gatwick Airport, with latitude and longitude coordinates of (-0.1821, 51.1536). Substituting the coordinates into the semi-versus formula, the distance between the current position of the flight and Gatwick Airport can be calculated. In this embodiment, the gap ratio threshold is... Set to 0.3.

[0093] For flight trajectory time series data, only the portion with a remaining arrival time of no more than 30 minutes is retained, i.e. The time interval was set to 30 minutes, and a sliding window method was used to construct a fixed-time series sample from this 30-minute data set. In the sample construction of this experiment, the length parameter of the sliding window... The time is fixed at 600 seconds, with a preset sliding step of 30 seconds. (Regarding length...) For sliding windows smaller than the length parameter, padding is added to the front of the sliding window sample. The fill value for each time step is set as follows in this embodiment: The sliding window samples are then standardized by ignoring padding values ​​and used as input to the LSTM model.

[0094] II. Model Training and Application

[0095] This experiment divides and sorts the dataset by date, taking all samples from the most recent 20% of dates as the validation set and all samples from the remaining 80% of dates as the training set. Weak and strong positive samples are constructed according to the method in Example 1. In the flight phase division rules, based on the fact that civil aviation instrument landing systems can typically provide stable glide path signals within a 10-nautical-mile range, and the glide angle is controlled between 2.5° and 3°, the distance of flight phase 2... and height The threshold values ​​were set at 10 nautical miles and 3,000 feet. Furthermore, according to relevant regulations and laws in many countries, the speed of civil aircraft is limited when flying below 10,000 feet. Combined with an ideal descent rate of 300 feet per minute, the distance for flight phase 1 in this experiment was... and height The threshold values ​​were set at 40 nautical miles and 10,000 feet. The remaining samples were assigned to flight phase 0.

[0096] In this experiment, the wind speed threshold for weak winds The value is 5 knots. When the wind speed in the meteorological data is lower than 5 knots, this wind speed is considered a weak wind and the meteorological phase is divided into phase 0. Only when the wind speed is higher than 5 knots will the meteorological phase be divided in detail according to the above introduction.

[0097] When constructing strong positive samples, the weights of all features can be determined by first using an LSTM network to perform flight arrival time prediction and then using the permutation importance method to determine the importance of each feature, thereby determining the weights.

[0098] III. Performance Evaluation Indicators and Methods

[0099] The basic idea of ​​permutation feature importance is to evaluate the impact of input features on model performance by randomly permuting them. The steps of this method are as follows: Original model evaluation: Calculate the model's performance metrics using the original data.

[0100] Feature permutation: For each feature, its value is randomly permuted, that is, all the values ​​of the feature are randomly shuffled, and only one feature is permuted at a time.

[0101] Model evaluation: The replaced data is used as input data for the model again, and the performance indicators of the current model are calculated.

[0102] Feature importance calculation: Calculate the differences between the various indicators of the model after the original data and the permuted data are used as inputs to the model. The greater the difference, the more important the corresponding feature.

[0103] To evaluate the performance of the flight prediction model trained in this experiment, flight trajectory features and meteorological features were used as model inputs based on the permutation importance results. The weights of each feature selected when calculating the PC-MTD value are as follows: Distance: 2.96, Azimuth Sine: 0.31, Azimuth Cosine: 0.97, Pressure Altitude: 3.38, Ground Speed: 1.15, Vertical Speed: 0.42, Ground Acceleration: 0.39, Glide Angle: 0.55, Track Angle Sine: 0.4, Track Angle Cosine: 0.73, Wind Speed: 0.15, Wind Direction Cosine: 0.38, Visibility: 0.36, Temperature-Dew Point Difference: 0.39.

[0104] Since features such as distance and air pressure altitude are more important to the model, this experiment uses a different formula for calculating the PC-MTD distance. middle, The trajectory feature weight is set to 0.9. The meteorological feature weight is set to 0.1. In this embodiment, The physical feature phase contrast learning loss accounts for 0.4% of the total loss function. The proportion of the PC-MTD contrastive learning loss in the total loss function is 0.6.

[0105] After the hybrid contrastive learning model has been trained, the encoder parameters at this point are used as the initialization parameters for the encoder in the regression phase, resulting in the intermediate representation vector of the samples. A fully connected layer is then added as the regression output layer for regression prediction training. The regression training uses the MAE loss function, and after training, a model for flight arrival time prediction is obtained.

[0106] IV. Experimental Data and Analysis

[0107] To verify the effectiveness of the proposed solution (denoted as PC-MTD) on real data, this experiment selected traditional contrastive learning models: Random Forest (RF), XGBoost, and AdaBoost; deep learning model: LSTM; and typical contrastive learning model: T-Loss as baseline models for comparison with the proposed solution. MAE and RMSE were selected as evaluation metrics for the models. Table 1 shows the comparison results of the evaluation metrics of the models in different solutions.

[0108] Table 1: Comparison of evaluation metrics between the PC-MTD model of this invention and various baseline models

[0109] As can be seen from the comparison results in Table 1, the model proposed in this invention (PC-MTD model) achieves the best performance in both MAE and RMSE indicators: MAE=2.157min and RMSE=3.334min. This indicates that the method can not only reduce the average prediction error, but also more effectively suppress the prediction error caused by noisy samples.

[0110] Compared to traditional machine learning models such as RF, XGBoost, and AdaBoost, the PC-MTD model of this invention further reduces both MAE and RMSE, indicating that traditional machine learning has limitations in characterizing the continuity and similarity of flight trajectories over time. In contrast, trajectory similarity-based representation learning can provide more refined and dynamically consistent feature expressions, thereby improving the accuracy of arrival time prediction.

[0111] Meanwhile, the PC-MTD model also outperforms the LSTM baseline, which directly performs supervised regression. This indicates that while deep learning alone can utilize sequence information, the representations learned by the model may not be robust enough under the influence of factors such as sample noise, differences in different flight stages, and distribution variations. By introducing contrastive learning, the encoder obtains more discriminative and generalizable temporal representations through the constraint that similar samples are closer and dissimilar samples are farther away, thereby further reducing the error.

[0112] It is worth noting that the error of the contrastive learning baseline T-Loss is significantly higher than that of other methods, reflecting that in this task, the lack of a similarity definition consistent with flight trajectory patterns and a high-quality sample pair construction strategy may introduce strong training noise and weaken the benefits of contrastive learning. In contrast, the PC-MTD model reduces the interference from pseudo-similar sample pairs by explicitly incorporating trajectory similarity into the positive and negative sample construction and representation learning process, resulting in more significant improvements in both MAE and RMSE evaluation metrics.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing comparative learning samples for flight arrival time prediction, characterized in that, It includes: Acquire historical arrival flight trajectory data and meteorological observation data for a designated airport; preprocess the acquired data to construct the feature data required for flight arrival time prediction. The trajectory data is classified by flight number and sorted by time. The tail segment of the trajectory data with remaining arrival time not exceeding a specified value is retained for modeling. A time series sample of a preset length is constructed using the sliding window method. Data in the window sample that is less than the preset length is pre-filled. The time series sample is then standardized by ignoring the filler values. Based on the trajectory features, meteorological features, and remaining arrival time in the end timestamp of any time series sample, the sample phase is divided, and each time series sample in the same phase is regarded as a weak positive sample. The feature combination distance between any two time series samples in the same phase is calculated. A specified number of nearest neighbors of each time series sample are selected as its strong positive samples, thus forming a global strong positive sample index matrix.

2. The method for constructing comparative learning samples for flight arrival time prediction as described in claim 1, characterized in that, Any two time series samples within the same phase i and j Feature combination distance between D ( i , j The formula for calculating ) is: ； In the above formula, express i and j The common effective time step set among them; express i and j In the t The feature combination distance at each time step; and They are respectively i and j In the t The difference in trajectory features and the difference in meteorological features at each time step; and These represent the weights of trajectory feature differences and meteorological feature differences in the feature combination distance, respectively.

3. The method for constructing comparative learning samples for flight arrival time prediction as described in claim 2, characterized in that, and The calculation formula is as follows: ； In the above formula, For the trajectory features, the first f dimensional weight vector, The first of the meteorological characteristics f A weight vector of dimension; express i and j In the f Feature difference in dimensional trajectory features; express i and j In the f The difference in characteristics between different meteorological features; F 1 and F 2 represents the dimensions of trajectory features and meteorological features, respectively.

4. The method for constructing comparative learning samples for flight arrival time prediction as described in claim 3, characterized in that, Methods for determining weak positive samples include: Extract any number i End timestamps of time series samples The distance in the corresponding trajectory features air pressure altitude The data, and the flight phases were determined. phase 1; Extract any number i End timestamps of time series samples Wind direction in the corresponding meteorological characteristics Wind speed Heading angle The data, and the delineation of meteorological phases. phase 2; Extract any number i End timestamps of time series samples The remaining arrival time is and define the arrival time phase. phase 3; The label for each time sample sequence is calculated using the following formula. label Based on the labels of each time sample sequence, any two time sample sequences with the same label are considered weak positive samples of each other. 。 5. The method for constructing comparative learning samples for flight arrival time prediction as described in claim 1, characterized in that, phase 1~ phase The method for dividing 3 is as follows: ； In the above formula, D1 and H1 are the distance threshold to the target airport and the flight pressure altitude threshold corresponding to phase 1, respectively, and D2 and H2 are the distance threshold to the target airport and the flight pressure altitude threshold corresponding to phase 2, respectively. ； In the above formula, W It is the weak wind speed threshold; and They represent the first i End timestamps of time series samples The wind speed components in the parallel and perpendicular directions of flight; phase The value 2 can be 0, 1, 2, or 3, representing weak wind, tailwind, headwind, and crosswind respectively. 。 6. The method for constructing comparative learning samples for flight arrival time prediction as described in claim 1, characterized in that: The trajectory data includes: timestamp, 24-bit ICAO address, latitude, longitude, ground speed, heading angle, vertical speed, flight call sign, whether on the ground, alarm status, special location marker, transponder code, barometric altitude, geometrical altitude, ADS-B transmitter serial number, and the hour period to which the data was collected. The meteorological observation data includes: observation time, airport ICAO code, wind direction, wind speed, visibility, cloud cover, cloud height, air temperature / dew point temperature, and surface air pressure.

7. The method for constructing comparative learning samples for flight arrival time prediction as described in claim 6, characterized in that: The preprocessing of trajectory data and meteorological observation data includes: (1) Convert the timestamps of the trajectory data to Coordinated Universal Time, delete the entire row of data with empty timestamps, and divide it into two columns according to timestamp and flight number to perform deduplication; (2) Delete the entire row of abnormal data in ground speed and barometric altitude; (3) The spherical distance from the flight path point to the target airport is calculated using the semi-sine formula based on the coordinates of the flight path point; (4) Perform second-by-second resampling and conditional interpolation on each data point for each flight; (5) Perform a 3-second window median filter on the continuous numerical fields within each flight to suppress spikes and jitter noise; (6) Resample the meteorological observation data minute by minute and then align it with the trajectory data second by second.

8. A method for predicting flight arrival time, characterized in that, It includes: A two-layer LSTM and a one-layer fully connected layer are used as the contrastive learning encoder; the contrastive learning encoder is used to learn from the input time sample sequence. X i Generate intermediate vector representation h i The dataset containing a large number of weak positive samples and strong positive sample labels, obtained by the contrastive learning sample construction method for flight arrival time prediction as described in any one of claims 1-7, is used to train the contrastive learning encoder using hybrid contrastive learning. In the hybrid contrastive learning training phase, the padding values ​​in the input time series samples are first masked, and then... h i A subsequent projection head and L2 normalization are performed to obtain embedding vectors specifically for contrastive learning. z i The hybrid contrastive training phase employs a loss function that simultaneously incorporates physical feature phase contrastive learning loss and feature combination distance contrastive learning loss. ; After the hybrid contrastive learning is completed, a fully connected layer is added after the contrastive learning encoder as a regression output layer to build a prediction model. The samples in the collected dataset are sorted by date, and all samples from the most recent 20% of dates are used as the validation set, with the remainder used as the training set. The MAE loss function is then applied. The flight arrival time prediction model is trained. The trained prediction model is used to predict flight arrival times based on the input time feature sequence.

9. The method for predicting flight arrival time as described in claim 8, characterized in that: Hybrid contrastive learning training includes phase contrastive learning based on physical features and distance contrastive learning based on feature combination; in the former, each time series sample in the dataset is considered a positive sample if its weak positive sample is positive and the rest are negative samples; in the latter, its strong positive sample is positive and the rest are negative samples. And / or, the loss function used in the hybrid contrastive learning training phase. for: ； in, and They are the first i The physical feature phase contrast learning loss and feature combination distance contrast learning loss are used for each time series sample. and They are and Weights in the total loss function; P label ( i ) indicates the first i A set of weak positive samples from a time series sample; P s ( i ) indicates the first i A set of strongly positive samples from a time series sample; z i and z j For the first i and the j Embedded vectors of time sample sequences; This is a scaling factor used to adjust the distribution of similarity between samples; And / or, the MAE loss function for: ； in, It is the first i A time series sample in The actual remaining arrival time of the hour. It is the predicted remaining arrival time output by the prediction model. M It is the number of samples in a training batch.

10. A flight arrival time prediction system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it creates a prediction model trained in the flight arrival time prediction method as described in claim 8 or 9, and outputs the predicted flight arrival time based on the input time series samples.

Images