An aircraft trajectory prediction method and system based on motion pattern prototype memory enhancement and residual gate memory network

Abstract

The application relates to an aircraft trajectory prediction method and system based on a motion pattern prototype memory enhancement and a residual gate memory network, and the method comprises the following steps: obtaining historical airport scene aircraft trajectory data and preprocessing the aircraft trajectory data to obtain an aircraft trajectory sequence; constructing a candidate trajectory sample from the aircraft trajectory sequence; further constructing a kinematic perception training sample set and a motion pattern enhancement feature sequence; inputting the motion pattern enhancement feature sequence into the residual gate memory network to generate a memory condition response; performing high-dimensional feature extraction on an intermediate hidden state to obtain a high-dimensional feature; performing prediction based on the high-dimensional feature to generate a future multi-step continuous coordinate trajectory prediction result; further updating the residual gate memory network until training convergence is achieved, so that a trained prediction model is obtained. Compared with the prior art, the application improves the prediction ability of the model for local nonlinear maneuvering trajectories.

Description

Technical Field

[0001] This invention relates to the fields of air traffic management, airport surface operation monitoring, trajectory prediction and intelligent transportation, and in particular to an aircraft trajectory prediction method and system based on motion pattern prototype memory enhancement and residual gated memory network. Background Technology

[0002] Airport surface aircraft trajectory prediction is a crucial foundation for intelligent airport surface management, taxiway path planning, operational situation awareness, and operational risk identification. As airport operational density increases, the operational status of aircraft near taxiways, runway thresholds, holding points, and intersections becomes more complex. Accurate prediction of future aircraft positions helps support surface traffic monitoring and operational decision-making.

[0003] Existing airport surface trajectory prediction methods typically model based on time-series information such as historical position, speed, or heading. One type of method uses traditional filtering or machine learning models to extrapolate the aircraft's future position; another type uses recurrent neural networks, attention networks, graph neural networks, or long sequence modeling networks to learn trajectory evolution patterns. While these methods can capture the relationship between historical and future trajectories to some extent, they still have the following problems in continuous coordinate trajectory prediction scenarios at airport surfaces: First, aircraft motion at airport surfaces exhibits a significant uneven distribution of motion states. Aircraft spend most of their time taxiing in straight lines or holding positions, while turning, crossing intersections, and other high-curvature maneuvers occur relatively infrequently. If a standard sliding window is used directly to construct training samples, the model training process is easily dominated by the majority of straight-line taxiing samples, thus weakening its ability to learn turning maneuvers.

[0004] Second, ADS-B trajectory data is typically recorded in the form of latitude and longitude coordinates and velocity components, and the absolute coordinate offsets between different airport areas or different taxiways are relatively large. If the raw latitude and longitude coordinates are directly input into the model, it is easy to introduce the problem of inconsistent coordinate scales, causing the model to focus too much on absolute geographical location rather than local motion changes.

[0005] Third, existing continuous trajectory prediction models typically perform linear mapping or superimpose fixed-position encoding on continuous features such as position and velocity, lacking a mechanism to dynamically match and inject corresponding pattern information from the set of learnable motion patterns based on the current local trajectory motion pattern. In local nonlinear maneuvering phases such as turning and deceleration, although different trajectory segments may have similar motion patterns, ordinary feature mapping is difficult to form a stable pattern-level representation, thus affecting the model's ability to identify and predict local maneuvering changes.

[0006] Fourth, while long-sequence prediction models are beneficial for maintaining long-term trajectory continuity, they may exhibit response lag at turns, decelerations, or local maneuver changes. Particularly in continuous coordinate regression tasks, the use of pure multiplicative gating or the lack of residual-preserving paths can lead to feature transmission attenuation, affecting the prediction accuracy of local maneuver trajectories.

[0007] The invention disclosed in CN121459644A presents an intelligent prediction method for aircraft taxiing trajectories that integrates spatiotemporal features and motion constraints. It utilizes a Long Short-Term Memory (LSTM) network to extract temporal features of the trajectory, employs a graph attention network to model the spatial topological relationships of the control path, and achieves effective integration of heterogeneous information through a feature fusion layer. A multi-component loss function integrating position, velocity, and acceleration is proposed to constrain the continuity and smoothness of the predicted trajectory. A lightweight data augmentation strategy and an adaptive residual connection mechanism are designed to improve the model's generalization ability and mitigate the accumulation of long-term prediction errors. However, this scheme cannot handle the problems of imbalanced motion state samples and inconsistent scales of continuous coordinates, resulting in low prediction accuracy.

[0008] Therefore, a technical solution is needed for predicting the continuous coordinate trajectory of aircraft at airport surfaces, which can simultaneously address the contradictions between imbalanced motion state samples, inconsistent continuous coordinate scales, and long-term trajectory consistency and local maneuver sensitivity. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the existing technology by providing an aircraft trajectory prediction method and system based on motion pattern prototype memory enhancement and residual gating memory network. This addresses the problems in the existing technology, such as ordinary sliding window training samples being easily dominated by straight taxiing states, scale inconsistencies in the original latitude and longitude inputs, prediction lag of long sequence models at local turning maneuvers, and the potential decay of continuous coordinate features due to pure multiplication gating.

[0010] The objective of this invention can be achieved through the following technical solutions: An aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network includes: The historical trajectory of the aircraft to be predicted is acquired and preprocessed; the preprocessed data is input into a pre-trained prediction model to obtain the future continuous coordinate trajectory of the aircraft; the training process of the prediction model includes: Historical airport surface aircraft trajectory data is acquired and preprocessed to obtain an aircraft trajectory sequence; candidate trajectory samples are constructed from the aircraft trajectory sequence based on a preset observation window length and a preset prediction window length; the candidate trajectory samples include historical observation windows and corresponding future prediction windows; The velocity components of the aircraft trajectory are obtained in the historical observation window, and a kinematic perception training sample set is constructed. A coordinate-unified input sequence is further constructed. Features are extracted from the coordinate-unified input sequence, and the extracted features are matched and enhanced using a learnable motion pattern prototype memory matrix to obtain a motion pattern enhanced feature sequence. The motion pattern enhancement feature sequence is input into the residual gated memory network, and a memory conditional response is generated by combining persistent memory information and contextual memory information. Based on the memory conditional response, an intermediate hidden state is generated. The memory readout result is further obtained, and the intermediate hidden state is modulated. The modulation result is added to the intermediate hidden state to obtain the residual gated output. High-dimensional features are further obtained. Prediction is performed based on the high-dimensional features, and the parameters of the residual gated memory network are updated until the training converges, resulting in a trained prediction model.

[0011] Furthermore, the residual gated memory network generates the memory conditional response based on the current input features, persistent memory information, and contextual memory information; the current input features and the memory conditional response are added together to obtain the intermediate hidden state; the intermediate hidden state is represented as: in, It is in a hidden state in the middle. For memory conditional response, Input features for the current time step. This indicates a normalization operation. Represents information in persistent memory. It represents contextual memory information, and Attn(·) represents the attention interaction function.

[0012] Furthermore, the intermediate hidden state is modulated based on the memory readout result; the modulation result is added to the intermediate hidden state to obtain the residual gated output; the high-dimensional feature is further obtained, and the corresponding acquisition expression is: in, High-dimensional features For a feedforward network, σ(·) represents the Sigmoid activation function, and ⊙ represents element-wise multiplication. Read the results of the updated memory. This indicates a normalization operation. For residual-gated output; U(⋅) represents neural memory update operation, R(⋅) represents neural memory read operation, S t-1 and S t These represent the neural memory states before and after the update, respectively.

[0013] Furthermore, based on the velocity components of the aircraft trajectory, the velocity modulus and cumulative heading angle change are calculated, and the candidate trajectory samples are divided into waiting state, straight taxiing state, and turning state to obtain the kinematic perception training sample set. At the same time, corresponding sliding step size and sample retention probability are set respectively, and differentiated sliding sampling and probability retention are performed on the candidate trajectory samples. When the maximum velocity modulus in the historical observation window is less than a preset velocity threshold, the corresponding candidate trajectory sample is classified as waiting state. When the maximum velocity modulus in the historical observation window is greater than or equal to the preset velocity threshold, and the cumulative heading angle change after masking is greater than a preset turning threshold, the corresponding candidate trajectory sample is classified as turning state. The remaining candidate trajectory samples are classified as straight taxiing state.

[0014] Furthermore, anchor point relative coordinate transformation and velocity scale transformation are performed on the historical observation window of the kinematic perception training sample set to obtain the coordinate-uniformed input sequence; feature extraction is further performed, and the extracted features are matched and enhanced using the motion pattern prototype memory enhancement module to obtain the motion pattern enhanced feature sequence. The corresponding matching and enhancement expressions include: in, Enhance feature sequences for motion patterns. Based on the hidden representation, W represents the matching distribution between input time-time features and motion pattern prototypes. u and W m For learnable projection matrices, To hide the feature dimension, W b M is a learnable projection matrix. pattern A prototype memory matrix for learnable movement patterns. To produce a learnable output projection matrix, For coordinate-uniform input sequences, express function.

[0015] Furthermore, the transformation expression corresponding to the anchor point relative coordinate transformation is: in, These are the position coordinates of the anchor point after relative coordinate transformation. For reference position, The longitude component is for the reference location. The latitude component is the reference position. Let be the position coordinates at time t. This is the location scale transformation factor; The transformation expression corresponding to the velocity scale transformation includes: in, The velocity components after velocity scale transformation. For velocity scale transformation factor, For the velocity component.

[0016] Furthermore, based on the meridional and zonal velocity components at each moment within the historical observation window, the change in heading angle between adjacent moments after the orbital correction is calculated. The corresponding calculation formulas include: in, This represents the change in heading angle between adjacent time points after angle correction. , These represent the heading angles at time t and time t-1, respectively.

[0017] Furthermore, based on the meridional and zonal velocity components at each moment in the historical observation window, the velocity modulus, heading angle, and the heading angle change at adjacent moments after meridian angle correction are calculated; and the heading angle changes below a preset velocity threshold are masked to obtain the masked cumulative heading angle change; the masked cumulative heading angle change is expressed as: , t = 2, …, W Among them, Γ i Let s represent the cumulative change in heading angle during the i-th historical observation window. t Let τ represent the velocity modulus at time t. s This represents the preset speed threshold, I(·) represents the indicator function, and W represents the length of the historical observation window. This represents the change in heading angle between adjacent time points after angle correction.

[0018] Furthermore, prediction is performed based on the high-dimensional features. The corresponding prediction process includes: performing affine mapping on the high-dimensional features at each observation time to obtain the corresponding future multi-step prediction vector; then aggregating the future multi-step prediction vectors at all observation times in terms of time dimension; reshaping the aggregated prediction vector into a matrix corresponding to the future prediction window length and the output coordinate dimension to obtain the prediction result of the future multi-step continuous coordinate trajectory; further calculating the loss function and updating the parameters of the residual gated memory network until training converges to obtain the trained prediction model.

[0019] The present invention also provides an aircraft trajectory prediction system based on motion pattern prototype memory enhancement and residual gated memory network, including a memory and a processor. The memory stores a computer program, and the processor calls the computer program to execute the steps of the method described above.

[0020] Compared with the prior art, the present invention has the following advantages: (1) This invention introduces a learnable motion pattern prototype memory matrix to match and enhance the features extracted from the coordinate-uniform input sequence, thereby enhancing the representation of local maneuver patterns. Furthermore, the constructed motion pattern enhanced feature sequence is input into the residual gated memory network, and memory conditional responses are generated by combining persistent memory information and contextual memory information. This enables the network to fuse the current motion state with long-term persistent memory and short-term contextual information, thereby maintaining long-term trajectory consistency using persistent memory and contextual memory. Furthermore, by using the memory readout results to modulate the intermediate hidden state and adding the modulation result to the intermediate hidden state, the residual gated output is obtained. While modulating the hidden state using the memory readout results, the original hidden state is preserved, reducing the risk of information decay caused by pure multiplication gate, thereby making the final trained prediction model more robust and accurate in long-term trajectory prediction. Motion pattern prototype memory enhancement is used to enhance the representation of local maneuver patterns, and the neural memory network is used to model long-term evolutionary relationships, thereby realizing the synergistic effect of local motion pattern enhancement and long-term trajectory memory modeling.

[0021] (2) This invention divides candidate trajectory samples into three states—waiting, straight taxiing, and turning—by calculating the velocity modulus and cumulative heading angle change, and sets corresponding sliding step size and sample retention probability for each state. Differentiated sliding sampling and probability retention are performed on the candidate trajectory samples, which can reduce the repetitive sampling density of stationary and straight motion samples and increase the effective sampling density of turning-related samples, thereby alleviating the problem of insufficient learning of turning maneuver samples caused by the dominance of straight taxiing samples in airport surface trajectories. The cumulative heading angle change after low-speed masking is used to identify the turning state, which can reduce the interference of speed direction jitter in waiting or low-speed taxiing states on turning discrimination, thereby effectively alleviating the model bias problem caused by the dominance of straight taxiing samples and the scarcity of key maneuver samples such as turning in airport surface data; thus ensuring the accuracy of the model in predicting unconventional maneuvers such as turning and waiting.

[0022] (3) This invention sets up a motion pattern prototype memory enhancement module before the neural memory prediction backbone. By constructing a memory matrix including multiple learnable motion pattern prototypes and calculating the matching weights between continuous trajectory hidden features and each motion pattern prototype, the corresponding motion pattern representations are dynamically aggregated and injected into the basic hidden features according to the matching results. This enables the model to explicitly enhance the maneuvering pattern information that matches the current local motion state before entering long-term neural memory modeling. Compared with directly using the original continuous coordinate features or fixed position encoding, this method can generate adaptive pattern enhancement representations based on the motion shape of the trajectory itself, thereby improving the representation ability of local nonlinear motion patterns such as turning and deceleration. The motion pattern enhancement features further work in synergy with dynamic long-term neural memory and residual gating to take into account both local maneuvering sensitivity and long-term trajectory consistency. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of the training method for an airport surface aircraft trajectory prediction model, which is based on motion pattern prototype memory enhancement and residual gated memory network, as provided in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the kinematic perception sample construction process of an aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the coordinate-unified input sequence construction for an aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the motion pattern memory enhancement module structure of an aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network provided in an embodiment of the present invention. Figure 5This is a schematic diagram of the residual gated memory interaction module structure of an aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of the direct multi-step trajectory output structure of an aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network provided in an embodiment of the present invention. Figure 7 This is a schematic diagram illustrating the execution flow of the airport surface aircraft trajectory prediction method provided in this embodiment of the invention during the training and prediction phases. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0025] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0026] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0027] Example 1 This embodiment provides an aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network. The method includes the following steps: The historical trajectory of the aircraft to be predicted is obtained and preprocessed; the preprocessed data is then input into a pre-trained prediction model to obtain the future continuous coordinate trajectory of the aircraft.

[0028] like Figure 1 As shown, the overall training process of the prediction model includes: acquiring airport surface ADS-B trajectory data, grouping and sorting trajectories, generating candidate windows, calculating velocity and heading, constructing kinematics-aware samples, inputting coordinate-uniformed data, enhancing motion pattern prototype memory, interacting with residual gating memory, directly outputting multi-step relative coordinates, outputting prediction results, calculating loss and updating model parameters, specifically including: S1: Acquire historical airport surface aircraft trajectory data and preprocess it to obtain an aircraft trajectory sequence; construct candidate trajectory samples from the aircraft trajectory sequence based on the preset observation window length and the preset prediction window length; the candidate trajectory samples include historical observation windows and corresponding future prediction windows; S101: Trajectory Data Acquisition and Preprocessing Acquire ADS-B trajectory data of aircraft at the airport surface. Each trajectory point includes at least the aircraft identifier, timestamp, longitude, latitude, meridional velocity component, and zonal velocity component. Group the trajectory points according to the aircraft identifier, and arrange them in ascending order of timestamp within each group to obtain several independent aircraft trajectory sequences.

[0029] The trajectory data is filtered for abnormal data, including deleting invalid timestamp trajectory points, removing trajectory sequences whose length is less than the sum of the preset observation window length and the preset prediction window length, and removing candidate trajectory samples with abnormal coordinate jumps in the future prediction window during the candidate sample generation process.

[0030] In one embodiment, the original ADS-B trajectory data is grouped according to the aircraft identifier, and candidate trajectory samples are generated in chronological order within the trajectory sequence corresponding to each aircraft identifier. The candidate trajectory samples are saved as independent data slices after being constructed by kinematic sensing samples, and are divided into training set, validation set and test set according to the data slices during the model training stage.

[0031] Specifically, such as Figure 7 As shown, after obtaining the trained model parameters, the historical ADS-B trajectory sequence of the aircraft to be predicted on the airport surface is obtained. Anchor point relative coordinate transformation and velocity scale transformation are performed on the historical ADS-B trajectory sequence to obtain the coordinate-uniformed input sequence. The coordinate-uniformed input sequence is then input into the trained residual gated memory network to output the prediction result of the future multi-step continuous coordinate trajectory.

[0032] S102: Candidate Trajectory Sample Generation Let the preset observation window length be W and the preset prediction window length be H. For each aircraft trajectory sequence, candidate trajectory samples are constructed using a sliding window approach. Each candidate trajectory sample includes a historical observation window X and a future prediction window Y, where the historical observation window X includes W consecutive historical state vectors, and the future prediction window Y includes H consecutive future position coordinates.

[0033] The historical state vector can be represented as: Where, p t lon and p t lbtLet v represent the longitude and latitude of the aircraft's position at time t, respectively. t lon and v t lbt Let represent the meridional velocity component and the zonal velocity component of the aircraft at time t, respectively.

[0034] S2: Obtain the velocity components of the aircraft trajectory in the historical observation window and construct a kinematic perception training sample set; further construct a coordinate-uniform input sequence; extract features from the coordinate-uniform input sequence, and use the learnable motion pattern prototype memory matrix to match and enhance the extracted features to obtain a motion pattern enhanced feature sequence; Specifically, The motion pattern prototype memory enhancement module performs basic feature mapping on the coordinate-uniformed input sequence, and uses multiple learnable motion pattern prototypes in the motion pattern prototype memory matrix for matching and weighted aggregation. The aggregated motion pattern representation is then injected into the basic hidden features to obtain the motion pattern enhanced feature sequence.

[0035] Specifically, The coordinate-uniform input sequence is constructed through anchor point relative coordinate transformation and velocity scale transformation, enabling the model to focus on learning local motion changes rather than absolute geographic coordinate offsets.

[0036] Preferred, Based on the meridional and zonal velocity components at each moment in the historical observation window, the velocity modulus, heading angle, and the heading angle change of adjacent moments after the meridional angle is corrected are calculated. The heading angle change at low speed moments is masked by a velocity threshold to obtain the cumulative heading angle change after low speed masking. The cumulative heading angle change after low speed masking is then used to mask the heading angle disturbance at low speed moments by a velocity threshold, so that velocity directional jitter in waiting or near-stationary states is not misjudged as real turning motion.

[0037] S201: Motion State Determination For each candidate trajectory sample, the velocity modulus is calculated based on the velocity components in the historical observation window: Calculate the heading angle based on the velocity components: To avoid errors in calculating the change in heading angle due to the angle crossing the boundary, an angle correction is performed on the heading angle difference between adjacent time points: To avoid misinterpreting speed and direction fluctuations during waiting or low-speed gliding as genuine turning motion, a speed threshold τ is used. sBy masking the low-speed moments, the cumulative change in heading angle due to low-speed masking is obtained for the i-th historical observation window: , t = 2, …, W Where I(·) is the indicator function. When s t ≥ τ s When s t < τ s When the heading angle changes, the corresponding change is not included in the cumulative value.

[0038] S202: Construction of Kinematic Perception Training Samples like Figure 2 As shown, based on the motion state of the candidate trajectory sample, differential sliding or differential preservation is performed on the candidate trajectory sample to obtain the kinematic perception training sample set.

[0039] Specifically, different sliding step sizes or retention probabilities can be set for the waiting state, straight-line gliding state, and turning state. Among them, the retention strength of turning state samples is higher than that of straight-line gliding state samples, and the retention strength of straight-line gliding state samples is higher than that of waiting state samples.

[0040] The candidate trajectory samples are further divided into waiting state, straight-line gliding state, and turning state. Based on the motion state of the candidate trajectory sample, corresponding sliding step size and sample retention probability are set for different motion states. The starting position of the next candidate trajectory window in the original trajectory sequence is controlled based on the sliding step size, and the inclusion of the current candidate trajectory sample in the training sample set is determined based on the sample retention probability. Specifically, a larger sliding step size and lower retention intensity are used for waiting state samples to reduce duplicate sampling between approximately stationary trajectory windows. Sparse sampling is performed on straight-line gliding state samples to reduce the dominant role of the numerous straight trajectories in the training process. A smaller sliding step size and increased retention intensity are used for turning state samples to increase the frequency of trajectory windows appearing in the training sample set at different stages of the turning process. The kinematic perception training sample set is composed of candidate trajectory samples after differentiated sliding sampling and probability retention.

[0041] In one alternative embodiment, the waiting state, the straight-line gliding state, and the turning state each employ different sliding step sizes; in another alternative embodiment, the waiting state, the straight-line gliding state, and the turning state each employ different retention probabilities; in yet another alternative embodiment, both differentiated sliding step sizes and differentiated retention probabilities are employed simultaneously.

[0042] Based on the velocity modulus and the cumulative change in heading angle due to low-speed masking, the candidate trajectory samples are divided into the following three categories: if the maximum velocity modulus within the historical observation window is less than the preset velocity threshold τ sIf the candidate trajectory sample is classified as a waiting state; if the maximum velocity magnitude within the historical observation window is greater than or equal to the preset velocity threshold τ s Furthermore, the cumulative change in heading angle during low-speed masking is Γ i If the value is greater than the preset steering threshold τθ, the candidate trajectory sample is classified as a turning state; the remaining candidate trajectory samples are classified as straight-line gliding states.

[0043] In this way, the kinematic perception training sample set can reduce the excessive dominance of straight-line gliding samples and waiting samples in the training process, and increase the effective occurrence frequency of turning maneuver samples in the training process.

[0044] Specifically, After classifying candidate trajectory samples into waiting, straight-line gliding, and turning states, the selection process is not simply based on category. Instead, motion state is used as a control condition for both the candidate window sliding process and the sample retention process. Specifically, a corresponding sliding step size and sample retention probability are set for each motion state. The sliding step size determines the starting position of the next candidate window in the original trajectory sequence after the current candidate window is processed, and the sample retention probability determines whether the current candidate window is written into the training sample set. This forms a joint processing procedure of "motion state discrimination - window start point adjustment - probability screening - training sample generation". In one specific embodiment, the sliding step sizes for the waiting, turning, and straight-line gliding states are set to W / 2, 1, and 2, respectively, and the corresponding sample retention probabilities are set to 0.20, 1.00, and 0.30, respectively. This allows highly repetitive waiting samples to be generated with low-density windows and retained with low intensity, while turning samples are generated with continuous windows and retained with high intensity. The motion state distribution of the training samples is adjusted from two dimensions: window generation density and candidate window retention intensity.

[0045] S203: Construction of Coordinate-Unified Input Sequence like Figure 3 As shown, the construction process of the coordinate-unified input sequence includes: For each historical observation window in the kinematic perception training sample set, a reference position within the historical observation window is selected as the anchor point. The reference position can be the end position, start position, middle position, or other preset position of the historical observation window.

[0046] Let the reference position be: The position coordinates p at time t in the historical observation window t Perform anchor point relative coordinate transformation: Where, λ pos This is the location scale transformation factor.

[0047] The velocity component v at time t in the historical observation window t Perform velocity scaling transformation: Where, λ vel This is the velocity scale transformation factor.

[0048] The relative position features and velocity features are concatenated to obtain the coordinate-consistent input features: The coordinate-uniform input sequence is composed of coordinate-uniform input features from all W time points: The aforementioned coordinate uniformity processing shifts the model's learning focus from absolute latitude and longitude to relative motion changes, thereby reducing the impact of absolute coordinate offsets in different airport areas or different taxiways on the prediction model.

[0049] S204: Motion Pattern Memory Enhancement To enhance the representation of turning, deceleration, and other local maneuvering patterns in continuous aircraft trajectories, this embodiment sets up a motion pattern prototype memory enhancement module before the neural memory prediction backbone. The module maintains a motion pattern prototype memory matrix composed of multiple learnable motion pattern prototypes and adaptively selects pattern information relevant to the current local motion state based on the matching relationship between the current trajectory hidden features and each motion pattern prototype.

[0050] like Figure 4 As shown, the process of enhancing motor pattern memory includes: Mapping the coordinate-unified input sequence X̃ to the hidden feature space yields the basic hidden representation H. base : Among them, W b It is a learnable projection matrix.

[0051] Set the motion pattern prototype memory enhancement module M pattern The motion pattern memory enhancement module maps the coordinate-uniformed input sequence to the hidden feature space, calculates the matching weights between the hidden features and the motion pattern memory matrix, and injects the motion pattern representation into the hidden features according to the matching weights, resulting in a motion pattern enhanced feature sequence. The motion pattern memory matrix is ​​used to store several motion pattern prototypes. The matching weights between the basic hidden representation and the motion pattern memory matrix are calculated as follows: Among them, W u and W mLet be the learnable projection matrix, d be the hidden feature dimension, and A represent the matching distribution between the features at input time step and the motion pattern prototype. The motion pattern representation is injected into the basic hidden representation according to the matching weights to obtain the motion pattern enhanced feature sequence: Among them, W c The output projection matrix is ​​learnable. Through this motion pattern memory enhancement process, the model can enhance the local maneuver pattern representation before entering the memory prediction backbone. In one embodiment, the motion pattern memory enhancement module is used to replace traditional absolute position encoding to avoid the absolute position encoding interfering with the consistency of continuous coordinate motion patterns.

[0052] The matching weights are obtained by calculating and normalizing the similarity between the query projection of the hidden features and the key projection of the motion pattern memory matrix; the motion pattern representation is obtained by weighted aggregation of the value projection of the motion pattern memory matrix.

[0053] Unlike directly using ordinary position encoding or only using linear feature mapping, this embodiment does not directly superimpose the preset temporal position vector onto the trajectory features. Instead, it dynamically generates a pattern enhancement term based on the matching relationship between the current continuous trajectory features and the learnable motion pattern prototype, so that trajectory segments with similar local motion patterns obtain similar pattern responses. Then, the response is input into the subsequent neural long-term memory network for temporal modeling.

[0054] S3: Input the motion pattern enhanced feature sequence into the residual gated memory network, and generate a memory conditional response by combining persistent memory information and contextual memory information; generate intermediate hidden states based on the memory conditional response; further obtain the memory readout results and modulate the intermediate hidden states; add the modulated results to the intermediate hidden states to obtain the residual gated output; further obtain high-dimensional features; make predictions based on the high-dimensional features and update the parameters of the residual gated memory network until training converges, and obtain the trained prediction model.

[0055] S301: Residual Gated Memory Interaction like Figure 5 As shown, the interaction process of residual gated memory interaction includes: The motion pattern enhanced feature sequence Z is input into the residual gated memory interaction module. Let z t This represents the motion perception feature at time t. The residual-gated memory interaction module first generates a memory conditional response based on the input features at the current time, persistent memory information, and contextual memory information.

[0056] Preferred, The residual gated memory network adopts a Titans-MAC-like neural memory backbone structure, which includes persistent memory path, contextual memory path, fast memory update path, and memory readout path. The persistent memory path is used to maintain the consistency of long-term trajectory, the contextual memory path is used to provide local motion context within the current observation window, the fast memory update path is used to update the memory state according to the intermediate hidden state, and the memory readout path is used to generate gated modulation signal.

[0057] This embodiment does not directly input the original continuous coordinate features into the neural memory backbone. Instead, it first obtains the motion pattern enhanced feature sequence through the motion pattern prototype memory enhancement module, and then uses this feature sequence as the continuous numerical input of the neural memory backbone, so that the local maneuver pattern representation is completed before the long-term temporal memory modeling. Subsequently, the long-term temporal modeling of the motion pattern enhanced features is performed using the persistent memory path, contextual memory path, and dynamic memory update path of the neural memory backbone. At the same time, an additive residual gated bypass is introduced in the memory readout stage, so that the memory readout result is used as a gated modulation signal to participate in the hidden state update, instead of directly suppressing the hidden state using a pure multiplication method.

[0058] Preferred, The motion pattern enhancement feature sequence is input into the residual gated memory network based on the Titans-MAC neural memory backbone, and memory conditional responses are generated using persistent memory paths and contextual memory paths. The gated modulation signal is obtained through memory update paths and memory readout paths.

[0059] In one embodiment, the memory conditional response can be represented as: Where Norm(·) represents the normalization operation. Represents information in persistent memory. Represents contextual memory information, Attn(·) represents the attention interaction function, and r t This indicates a memory-conditional response.

[0060] The intermediate hidden state is obtained by adding the memory conditional response to the current input feature: Subsequently, based on the updated memory reading results Generate a gated signal and obtain a residual gated output through additive residual gating: Where σ(·) represents the Sigmoid activation function, and ⊙ represents element-wise multiplication. For residual-gated output; U(⋅) represents neural memory update operation, R(⋅) represents neural memory read operation, S t-1 and S t These represent the neural memory states before and after the update, respectively.

[0061] Compared to pure multiplicative gating, the additive residual gating method described above modulates the hidden state by utilizing the memory readout result while preserving the intermediate hidden state itself, which helps reduce the risk of information decay in continuous coordinate regression tasks.

[0062] Furthermore, a feedforward residual transformation can be performed on the residual-gated output to obtain high-dimensional features: Where FFN(·) represents a feedforward network, u t This represents the refined feature at time t.

[0063] Specifically, In terms of model structure, the residual-gated memory network does not simply perform the addition of hidden states, but sets up a motion pattern prototype memory enhancement process before the residual-gated memory interaction. Specifically, the coordinate-unified trajectory features are first mapped to the hidden feature space, and matching weights are calculated with multiple motion pattern prototypes in the motion pattern prototype memory enhancement module. The motion pattern representations are weighted and aggregated according to the matching weights and injected into the basic hidden features to obtain the motion pattern enhanced feature sequence. Subsequently, the motion pattern enhanced feature sequence is input into the Titans-MAC-like dynamic neural memory backbone. After forming the current intermediate hidden state, the neural memory state is updated, and the memory readout result is re-obtained from the updated neural memory state. The memory readout result generates a gating signal to perform additive residual modulation on the current hidden state. Furthermore, in the trajectory output stage, instead of using only the hidden features of the last moment of the historical window for prediction, the high-dimensional features of each moment within the historical observation window are used to generate multi-step prediction vectors corresponding to the complete future prediction window through affine mapping. Then, the multi-step prediction vectors are aggregated along the historical time dimension and reconstructed to obtain the final future multi-step continuous coordinate trajectory. This forms a continuous processing process of "motion pattern prototype matching enhancement - dynamic neural memory update - updated memory readout gate control - full historical window multi-step prediction aggregation".

[0064] S302: Direct multi-step trajectory output like Figure 6 As shown, the process of direct multi-step trajectory output includes: Perform an affine mapping on the refined feature at each observation time to obtain the corresponding multi-step prediction vector for the future: Among them, W o and b o Let q represent the weights and biases of the affine mapping, respectively. t This represents the multi-step prediction vector for the future generated from the features at the t-th observation time.

[0065] The future multi-step prediction vectors for all observation times are aggregated along the time dimension. In one embodiment, time mean pooling is used: The aggregated prediction vector is reshaped into a future prediction window length H and an output coordinate dimension C. out The corresponding matrix yields the prediction results of the continuous coordinate trajectory over multiple future steps: This output method directly generates the complete future trajectory without recursively inputting the prediction results from the previous step into the next prediction process, thus reducing the risk of error propagation in autoregressive prediction.

[0066] S303: Model Training A loss function is constructed based on the error between the predicted future multi-step continuous coordinate trajectory Ŷ and the actual future trajectory Y, and the parameters of the trajectory prediction model are updated.

[0067] In an alternative embodiment, the Huber loss function or the smoothed Huber loss function is used as the training objective to reduce the impact of abnormal coordinate bias on the model training process.

[0068] In an alternative embodiment, the AdamW optimizer is used to update the model parameters, and at least one of gradient pruning, learning rate scheduling, and learning rate warm-up strategies is employed to stabilize the training.

[0069] In one specific embodiment, the observation window length W is 30, the prediction window length H is 15, and the turning threshold τ is... θ Take 15°, velocity threshold τ s Set to 5.0; different sliding step sizes and retention probabilities are used for the waiting state, turning state, and straight-line coasting state; position scale transformation factor λ. pos Set to 1000, velocity scale transformation factor λ vel The value is set to 0.01; the motion pattern memory matrix includes 32 motion pattern memory slots; the hidden feature dimension is set to 256. The above parameters are only one embodiment and do not constitute a limitation on the scope of protection of this invention.

[0070] Example 2 This embodiment illustrates how the residual gated memory network trained in Embodiment 1 is used in the prediction phase.

[0071] Obtain the historical ADS-B trajectory sequence of the aircraft to be predicted on the airport surface. This historical ADS-B trajectory sequence includes the position coordinates and velocity components at multiple consecutive historical moments.

[0072] Anchor point relative coordinate transformation and velocity scale transformation are performed on the historical ADS-B trajectory sequence to obtain a coordinate-uniform input sequence.

[0073] The coordinate-unified input sequence is input into the residual gated memory network trained by Example 1.

[0074] The residual gated memory network first matches the input features with the motion pattern prototype memory enhancement module through the motion pattern memory enhancement module, and injects the motion pattern representation into the input features; then, it performs long-term memory interaction and residual gated modulation on the motion perception features through the residual gated memory interaction module; finally, it generates the prediction results of future multi-step continuous coordinate trajectory through affine mapping and temporal dimension aggregation.

[0075] This prediction process can be used for airport surface operations monitoring, taxiing situation prediction, assisting in scheduling decisions, or other applications that require predicting the future position of aircraft.

[0076] Example 3 This embodiment demonstrates the technical effectiveness of the method described in Embodiment 1, and a comparative experiment is conducted on the same airport surface ADS-B trajectory dataset. In the experiment, the historical observation window length... Set to 30, future prediction window length Set to 15. Evaluation metrics include Average Displacement Error (ADE), Endpoint Displacement Error (FDE), Turn Sample Average Displacement Error (Turn ADE), and Turn Sample Endpoint Displacement Error (Turn FDE). ADE measures the average deviation of the predicted trajectory over the entire prediction window, FDE measures the deviation between the predicted endpoint and the actual endpoint, and Turn ADE and Turn FDE measure the prediction error on turn-related samples.

[0077] in, Indicates the number of test samples. Indicates the prediction window length. Indicates the first The sample at the th The predicted position for each prediction step. This indicates the corresponding actual position. Turn ADE and Turn FDE are the ADE and FDE calculated on the subset of turn-related samples.

[0078] Comparative Example 1: Natural Sampling Training Method Without Using Kinematic Perception Sample Construction Comparative Example 1 uses the same residual-gated memory network backbone as this invention, but its training window is constructed using a natural sliding sampling method, without performing differentiated sample retention based on waiting, straight-line sliding, and turning states. The method of this invention, however, uses the aforementioned kinematic perception sample construction method to perform differentiated sliding or differentiated retention on candidate trajectory samples for different motion states. Both methods are evaluated on the same natural test window, which does not perform motion state-related rebalancing.

[0079] The experimental results are shown in Table 1.

[0080] Table 1. Impact of kinematic perception sample construction on prediction performance As shown in Table 1, under the same natural test window, the method of this invention achieves lower errors than Comparative Example 1 in terms of overall average displacement error, overall endpoint displacement error, average displacement error of turning samples, and endpoint displacement error of turning samples. This result indicates that the kinematic perception sample construction can improve the model's ability to learn local nonlinear motion segments of the airport surface. This effect is not caused by test window rebalancing, but rather by the differentiated construction of samples for different motion states during the training phase.

[0081] Comparative Examples 2 to 4: Ablation Comparison of Different Model Structural Variants To further verify the role of the residual gating memory interaction structure and motion pattern memory enhancement module in this invention, the following structural variants are set as comparative examples.

[0082] Comparative Example 2 is a linear_residual variant, which uses linear residual connections instead of the residual gated memory interaction method in this invention.

[0083] Comparative Example 3 is a pure_mul variant, which uses pure multiplication gating instead of the additive residual gating method in this invention, that is, it does not retain the additive residual path of the intermediate hidden state.

[0084] Comparative Example 4 is a PE variant, which uses absolute position encoding instead of the motion pattern memory enhancement method in this invention.

[0085] Table 2. Comparison of ablation results for different model structural variants As shown in Table 2, the method of this invention achieves lower values ​​in both overall prediction error and turn-related prediction error compared to the linear residual variant, the pure multiplicative gating variant, and the absolute position coding variant (PE). In particular, the pure_mul variant shows a significant increase in both the overall average displacement error and the overall endpoint displacement error, indicating that in continuous coordinate trajectory regression tasks, using only pure multiplicative gating can easily lead to feature transmission attenuation or optimization instability. In contrast, this invention employs additive residual gating, which modulates intermediate hidden states using memory readout results while preserving the original hidden states, thus improving the stability of continuous coordinate prediction.

[0086] Meanwhile, the PE variant outperforms the method of this invention in both the average displacement error (Turn ADE) and the final displacement error (TurnFDE) of the turning sample, indicating that the motion pattern memory enhancement module can more effectively enhance the representation of local maneuver patterns, thereby improving the prediction performance on turning-related samples.

[0087] Comparative Example 5: Performance Comparison of Different Prediction Methods To illustrate the overall performance of the method of this invention in airport surface aircraft trajectory prediction, it is compared with methods such as Kalman filtering, PatchTST, Mamba, iTransformer, Transformer, and XGBoost. All methods use the same airport surface ADS-B trajectory dataset and are tested under the same evaluation metrics. The experimental results are shown in Table 3.

[0088] Table 3 Performance Comparison Results of Different Prediction Methods As shown in Table 3, the method of this invention achieves lower errors in the Overall Average Displacement Error (ADE), Overall Final Displacement Error (FDE), and Turn ADE metrics, indicating that the method can achieve good results in terms of overall trajectory prediction accuracy and average prediction accuracy of turning-related samples. In terms of the Turn FDE metric, the method of this invention achieves lower errors compared to Kalman, PatchTST, Mamba, iTransformer, and XGBoost, but higher than Transformer. Therefore, the technical effect of the method of this invention is mainly reflected in improving overall prediction accuracy and average trajectory prediction accuracy during turning, rather than achieving absolute superiority in all final displacement error metrics.

[0089] As can be seen from Tables 1 to 3, this invention, through the combination of kinematic perception sample construction, coordinate-uniform input, motion pattern memory enhancement, and additive residual gated memory interaction, can alleviate the problem of insufficient learning of turning maneuvers caused by the dominance of straight-line taxiing state in airport surface trajectory samples, reduce feature scale interference in continuous coordinate trajectory prediction, and improve the prediction performance of the model in local nonlinear maneuver scenarios.

[0090] The above comparative examples and the method of this invention use the same data partitioning, the same input window length, and the same prediction window length, and are tested under the same evaluation index. The experimental results are shown in Table 2.

[0091] Example 4 This embodiment further illustrates the collaborative processing of the motion pattern prototype memory enhancement module and the residual gated neural memory network.

[0092] Obtain the coordinate-uniformed input sequence constructed according to Example 1, and input it into the motion pattern prototype memory enhancement module. Match the basic hidden features with multiple learnable motion pattern prototypes in the motion pattern prototype memory matrix, perform weighted aggregation on the corresponding motion pattern representations according to the matching weights, and inject the aggregation result into the basic hidden features to obtain the motion pattern enhancement feature sequence.

[0093] The motion pattern enhanced feature sequence is input into the residual gated neural memory network. Temporal modeling is performed through persistent memory path, contextual memory path and memory update path. The updated memory readout results are used to perform additive residual gate modulation on the intermediate hidden state to obtain high-dimensional features.

[0094] Finally, multi-step prediction vectors for future steps are generated from the high-dimensional features at each moment within the historical observation window, and the prediction results of the aircraft's future multi-step continuous coordinate trajectory are obtained by aggregating them through the time dimension.

[0095] In this way, the motion pattern prototype memory enhancement module is used to enhance the representation of local maneuver patterns, and the neural memory network is used to model the long-term evolution relationship of motion pattern enhancement features, thereby realizing the synergistic connection between local motion pattern enhancement and long-term trajectory memory modeling.

[0096] Example 5 This embodiment provides an aircraft trajectory prediction system based on motion pattern prototype memory enhancement and residual gated memory network, including a memory and a processor. The memory stores a computer program, and the processor calls the computer program to execute the steps of the method as described in Embodiment 1.

[0097] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. An aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network, characterized in that, include: Acquire and preprocess the historical trajectory of the aircraft to be predicted; The preprocessed data is input into a pre-trained prediction model to obtain the aircraft's future continuous coordinate trajectory. The training process of the prediction model includes: Historical airport surface aircraft trajectory data is acquired and preprocessed to obtain an aircraft trajectory sequence; candidate trajectory samples are constructed from the aircraft trajectory sequence based on a preset observation window length and a preset prediction window length; the candidate trajectory samples include historical observation windows and corresponding future prediction windows; The velocity components of the aircraft trajectory are obtained in the historical observation window, and a kinematic perception training sample set is constructed. A coordinate-unified input sequence is further constructed. Features are extracted from the coordinate-unified input sequence, and the extracted features are matched and enhanced using a learnable motion pattern prototype memory matrix to obtain a motion pattern enhanced feature sequence. The motion pattern enhancement feature sequence is input into the residual gated memory network, and a memory conditional response is generated by combining persistent memory information and contextual memory information. Based on the memory conditional response, an intermediate hidden state is generated. The memory readout result is further obtained, and the intermediate hidden state is modulated. The modulation result is added to the intermediate hidden state to obtain the residual gated output. High-dimensional features are further obtained. Prediction is performed based on the high-dimensional features, and the parameters of the residual gated memory network are updated until the training converges, resulting in a trained prediction model.

2. The aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network according to claim 1, characterized in that, The residual gated memory network generates the memory conditional response based on the current input features, persistent memory information, and contextual memory information; the current input features and the memory conditional response are added together to obtain the intermediate hidden state; the intermediate hidden state is represented as follows: in, It is in a hidden state in the middle. For memory conditional response, Input features for the current time step. This indicates a normalization operation. Represents information in persistent memory. It represents contextual memory information, and Attn(·) represents the attention interaction function.

3. The aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network according to claim 1, characterized in that, The intermediate hidden state is modulated based on the memory readout result; the modulation result is added to the intermediate hidden state to obtain the residual gated output; the high-dimensional feature is further obtained, and the corresponding acquisition expression is: in, High-dimensional features For a feedforward network, σ(·) represents the Sigmoid activation function, and ⊙ represents element-wise multiplication. Read the results of the updated memory. This indicates a normalization operation. For residual-gated output; U(⋅) represents neural memory update operation, R(⋅) represents neural memory read operation, S t-1 and S t These represent the neural memory states before and after the update, respectively.

4. The aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network according to claim 1, characterized in that, Based on the velocity components of the aircraft trajectory, the velocity modulus and cumulative heading angle change are calculated, and the candidate trajectory samples are divided into waiting state, straight taxiing state and turning state to obtain the kinematic perception training sample set; at the same time, the corresponding sliding step size and sample retention probability are set respectively, and differential sliding sampling and probability retention are performed on the candidate trajectory samples. When the maximum velocity magnitude within the historical observation window is less than the preset velocity threshold, the corresponding candidate trajectory sample is classified as a waiting state; when the maximum velocity magnitude within the historical observation window is greater than or equal to the preset velocity threshold, and the cumulative change in heading angle after masking is greater than the preset turning threshold, the corresponding candidate trajectory sample is classified as a turning state; the remaining candidate trajectory samples are classified as straight-line taxiing states.

5. The aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network according to claim 1, characterized in that, Anchor point relative coordinate transformation and velocity scale transformation are performed on the historical observation window of the kinematic perception training sample set to obtain the coordinate-uniformed input sequence; further feature extraction is performed, and the extracted features are matched and enhanced using the motion pattern prototype memory enhancement module to obtain the motion pattern enhanced feature sequence. The corresponding matching and enhancement expressions include: in, Enhance feature sequences for motion patterns. Based on the hidden representation, W represents the matching distribution between input time-time features and motion pattern prototypes. u and W m For learnable projection matrices, To hide the feature dimension, W b M is a learnable projection matrix. pattern A prototype memory matrix for learnable movement patterns. To produce a learnable output projection matrix, For coordinate-uniform input sequences, express function.

6. The aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network according to claim 5, characterized in that, The transformation expression corresponding to the relative coordinate transformation of the anchor point is: in, These are the position coordinates of the anchor point after relative coordinate transformation. For reference position, The longitude component is for the reference location. The latitude component is the reference position. Let be the position coordinates at time t. This is the location scale transformation factor; The transformation expression corresponding to the velocity scale transformation includes: in, The velocity components after velocity scale transformation. For velocity scale transformation factor, For the velocity component.

7. The aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network according to claim 1, characterized in that, The change in heading angle between adjacent moments after correction is calculated based on the meridional and zonal velocity components at each moment in the historical observation window, and the corresponding calculation formula is as follows. include: in, This represents the change in heading angle between adjacent time points after angle correction. , These represent the heading angles at time t and time t-1, respectively.

8. The aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network according to claim 7, characterized in that, The velocity modulus, heading angle, and the change in heading angle between adjacent moments are calculated based on the meridional and zonal velocity components at each moment in the historical observation window. The changes in heading angle below a preset speed threshold are masked to obtain the masked cumulative heading angle change; the masked cumulative heading angle change is expressed as: ,t = 2, …, W Among them, Γ i Let s represent the cumulative change in heading angle during the i-th historical observation window. t Let τ represent the velocity modulus at time t. s This represents the preset speed threshold, I(·) represents the indicator function, and W represents the length of the historical observation window. This represents the change in heading angle between adjacent moments after angle correction.

9. The aircraft trajectory prediction method based on motion pattern prototype memory enhancement and residual gated memory network according to claim 1, characterized in that, The prediction process based on the high-dimensional features includes: performing an affine mapping on the high-dimensional features at each observation time to obtain the corresponding future multi-step prediction vector; then aggregating the future multi-step prediction vectors at all observation times in terms of time dimension; reshaping the aggregated prediction vectors into a matrix corresponding to the future prediction window length and the output coordinate dimension to obtain the prediction result of the future multi-step continuous coordinate trajectory; further calculating the loss function and updating the parameters of the residual gated memory network until training converges to obtain the trained prediction model.

10. An aircraft trajectory prediction system based on motion pattern prototype memory enhancement and residual gated memory network, characterized in that, It includes a memory and a processor, the memory storing a computer program, the processor invoking the computer program to perform the steps of the method as described in any one of claims 1 to 9.

Images