Trajectory prediction method and related device, electronic device and storage medium

Abstract

The application discloses a flight path prediction method and related device, electronic equipment and storage medium, wherein the flight path prediction method comprises: performing feature embedding based on attribute information of an observation track point to obtain embedded feature representation; performing semantic extraction based on the embedded feature representation to obtain semantic feature representation; performing route prediction based on the semantic feature representation to obtain route probability of each expert network, and based on the route probability of each expert network and output feature representation processed by each expert network respectively, obtaining target feature representation, and performing mapping prediction based on the target feature representation to obtain attribute value of a flight attribute; obtaining attribute information of a predicted track point based on the attribute value of each flight attribute in the predicted track point; wherein the predicted track point is located at a later moment of the observation track point. The above scheme can improve the accuracy of flight path prediction.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a trajectory prediction method and related devices, electronic devices and storage media. Background Technology

[0002] Aircraft such as drones, electric vertical take-off and landing aircraft, and small general aviation vehicles have been widely used in urban logistics, surveying and inspection, and emergency rescue. With the surge in the number of aircraft and the diversification of their operating status, trajectory prediction tasks are facing more severe challenges.

[0003] In actual flight, an aircraft is described by several flight attributes, which not only constitute the basic representation of its motion state at the single-target level but also exhibit typical correlation and difference characteristics at the multi-target cooperative level. Currently, existing technologies typically encode and convert the attribute information of track points before using it as model input, and model the prediction problem as a multi-classification problem concerning various flight attributes to perform autoregressive predictions to obtain the attribute information of track points at subsequent time points. However, this approach cannot fully utilize the inherent characteristics of different flight attributes, thus limiting the accuracy of track prediction. Therefore, improving the accuracy of track prediction has become an urgent problem to be solved. Summary of the Invention

[0004] The main technical problem addressed by this application is to provide a trajectory prediction method and related devices, electronic equipment, and storage medium that can improve the accuracy of trajectory prediction.

[0005] To address the aforementioned technical problems, the first aspect of this application provides a trajectory prediction method, comprising: embedding features based on attribute information of observed trajectory points to obtain embedded feature representations; wherein the attribute information includes attribute values ​​of several flight attributes; performing semantic extraction based on the embedded feature representations to obtain semantic feature representations; for each flight attribute: performing route prediction based on the semantic feature representations to obtain the routing probabilities of each expert network in an expert network group applicable to the flight attribute, and obtaining a target feature representation of the flight attribute based on the routing probabilities of each expert network and the output feature representations processed by each expert network respectively, and performing mapping prediction based on the target feature representations to obtain the attribute values ​​of the flight attributes in the predicted trajectory point; wherein the expert network group applicable to the flight attribute includes: a shared expert network applicable to all flight attributes and a customized expert network specifically for flight attributes; obtaining attribute information of the predicted trajectory point based on the attribute values ​​of each flight attribute in the predicted trajectory point; wherein the predicted trajectory point is located after the observed trajectory point.

[0006] To address the aforementioned technical problems, a second aspect of this application provides a trajectory prediction device, comprising: a feature embedding module, a semantic extraction module, an attribute prediction module, and a trajectory acquisition module. The feature embedding module is used to embed features based on attribute information of observed trajectory points to obtain embedded feature representations; wherein the attribute information includes attribute values ​​of several flight attributes. The semantic extraction module is used to extract semantics based on the embedded feature representations to obtain semantic feature representations. The attribute prediction module is used for each flight attribute to: perform route prediction based on the semantic feature representations to obtain the routing probabilities of each expert network in an expert network group applicable to the flight attribute; and obtain a target feature representation of the flight attribute based on the routing probabilities of each expert network and the output feature representations processed by each expert network according to the semantic feature representations; and perform mapping prediction based on the target feature representations to obtain the attribute values ​​of the flight attributes in the predicted trajectory point. The expert network group applicable to the flight attribute includes: a shared expert network applicable to all flight attributes and a customized expert network specifically for the flight attribute. The trajectory acquisition module is used to obtain attribute information of the predicted trajectory point based on the attribute values ​​of each flight attribute in the predicted trajectory point; wherein the predicted trajectory point is located after the observed trajectory point.

[0007] To address the aforementioned technical problems, a third aspect of this application provides an electronic device comprising at least a memory and a processor coupled to each other, wherein the memory stores at least program instructions, and the processor executes the program instructions to implement the trajectory prediction method of the first aspect.

[0008] To address the aforementioned technical problems, a fourth aspect of this application provides a computer-readable storage medium storing program instructions executable by a processor, the program instructions being used to implement the trajectory prediction method of the first aspect described above.

[0009] The above scheme embeds features based on the attribute information of observed waypoints to obtain embedded feature representations. The attribute information includes the attribute values ​​of several flight attributes. Semantic extraction is then performed based on the embedded feature representations to obtain semantic feature representations. For each flight attribute: route prediction is performed based on the semantic feature representations to obtain the routing probabilities of each expert network in the expert network group applicable to the flight attribute. Based on the routing probabilities of each expert network and the output feature representations processed by each expert network, the target feature representation of the flight attribute is obtained. Mapping prediction is then performed based on the target feature representations to obtain the attribute values ​​of the flight attributes in the predicted waypoints. The expert network group applicable to the flight attributes includes a shared expert network applicable to all flight attributes and a customized expert network specifically for flight attributes. Finally, based on the attribute values ​​of each flight attribute in the predicted waypoints, the attribute information of the predicted waypoints is obtained. Since the predicted waypoints are located after the observed waypoints, the path prediction... Instead of modeling it as a multi-class classification problem for each flight attribute, this approach predicts each flight attribute separately. Firstly, it performs route prediction based on semantic feature representations for each flight attribute, which helps to fully consider the unique characteristics of different flight attributes. Secondly, when predicting the attribute values ​​of each flight attribute, it relies on an expert network group containing a shared expert network and a custom expert network. This allows the custom expert network to fully learn the unique characteristics of different flight attributes, and the shared expert network to further learn the general temporal features of flight trajectories. Finally, using the routing probabilities of each expert network in the expert network group applicable to the flight attribute obtained during the route prediction process, as well as the output feature representations processed by each expert network, the target feature representation of the flight attribute is obtained. This approach avoids the predicament of different flight attributes being forced to share the same feature representation for attribute prediction, thus improving the accuracy of mapping and predicting the attribute values ​​of flight attributes. Therefore, it can improve the accuracy of trajectory prediction. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an embodiment of the trajectory prediction method of this application; Figure 2 This is a schematic diagram of a process of an embodiment of the trajectory prediction method of this application; Figure 3 This is a schematic diagram of the framework of an embodiment of the trajectory prediction device of this application; Figure 4 This is a schematic diagram of the framework of an embodiment of the electronic device of this application; Figure 5 This is a schematic diagram of a framework of an embodiment of the computer-readable storage medium of this application. Detailed Implementation

[0011] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0012] In the following description, specific details such as particular system architectures, interfaces, and technologies are presented for illustrative purposes rather than for limiting purposes, in order to provide a thorough understanding of this application.

[0013] In this paper, the terms "system" and "network" are often used interchangeably. The term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the slash " / " generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, "many" in this paper indicates two or more objects.

[0014] Please see Figure 1 , Figure 1 This is a schematic flowchart of an embodiment of the trajectory prediction method of this application. It should be noted that the process operations in this embodiment can be performed by an electronic device with computing capabilities or related equipment containing electronic devices. The specific structure and type of the electronic device and related equipment containing electronic devices are not limited herein. Specifically, this embodiment may include the following steps: Step S11: Based on the attribute information of the observed track points, perform feature embedding to obtain the embedded feature representation.

[0015] In this embodiment of the disclosure, the attribute information includes attribute values ​​of several flight attributes. Exemplarily, these flight attributes may include, but are not limited to, longitude, latitude, altitude, heading, speed, country of origin, and aircraft type; no limitation is made to the flight attributes herein. Furthermore, the observed waypoint represents the actual waypoint of the aircraft during flight, while the subsequently predicted waypoint represents the waypoint predicted for a subsequent moment (e.g., the next moment) based on the observed waypoint. Exemplarily, the predicted waypoint can be used to guide or control the aircraft's flight. Of course, the observed waypoint and the predicted waypoint can belong to the same aircraft.

[0016] In one implementation scenario, as a possible approach, the attribute values ​​of each flight attribute in the observed track points can be discretized to generate high-dimensional real-valued embedding vectors. These embedding vectors are then concatenated to form the joint representation vector of the observed track points, i.e., the embedding feature representation. For ease of description, the observed track point at time t can be denoted as X. t Several flight attributes can be denoted as at ∈ {lat,lon,sog,cog,coun,type,height}, then the discretized attribute values ​​are... It can be represented as: (1) Based on this, the embedding vectors of each flight attribute It can be represented as: (2) By concatenating the embedding vectors of the various flight attributes mentioned above, we can obtain the embedded feature representation e. t : (3) In another implementation scenario, as another possible approach, based on whether different flight attributes are correlated, several flight attributes can be divided into several attribute combinations (i.e., there is correlation between flight attributes within the same attribute combination, and no correlation between flight attributes in different attribute combinations). Then, the attribute values ​​of each flight attribute within the same attribute combination are embedded and mapped using the same embedding matrix to obtain the embedded representation of the corresponding attribute combination. Furthermore, the embedded feature representation can be obtained by fusing the embedding representations of several attribute combinations. For example, taking several flight attributes that may include longitude, latitude, altitude, heading, speed, country, and aircraft type as an example, the longitude and latitude attribute values ​​in the attribute information of the observed track point can be embedded and mapped based on a first embedding matrix to obtain a first embedding representation. The country and aircraft type attribute values ​​in the attribute information of the observed track point can be embedded and mapped based on a second embedding matrix to obtain a second embedding representation. The heading and speed attribute values ​​in the attribute information of the observed track point can be embedded and mapped based on a third embedding matrix to obtain a third embedding representation. Finally, the altitude attribute value in the attribute information of the observed track point can be embedded and mapped based on a fourth embedding matrix to obtain a fourth embedding representation. Based on this, the first, second, third, and fourth embedding representations can be fused to obtain the embedded feature representation. It should be noted that the number of rows in the first embedding matrix can be the sum of the sizes of the longitude and latitude vocabularies, and the number of columns in the first embedding matrix can represent the dimension of the joint longitude and latitude embedding. The number of rows in the second embedding matrix can be the sum of the sizes of the country and aircraft type vocabularies, and the number of columns in the second embedding matrix can represent the dimension of the joint country and aircraft type embedding. The number of rows in the third embedding matrix can be the sum of the sizes of the airspeed and heading vocabularies, and the number of columns in the third embedding matrix can represent the dimension of the joint airspeed and heading embedding. For ease of understanding, the first embedding representation is lon_lat. emb It can be represented as: (4) In the above formula (4), Embeding represents (lon_size + lat_size). D llThe first embedding matrix is ​​of size, where lon_size represents the size of the longitude vocabulary (e.g., it can exhaustively list all possible longitudes of an aircraft within a certain flight range expressed in "degrees, minutes, and seconds," and its total number can be considered the size of the longitude vocabulary), and lat_size represents the size of the latitude vocabulary (e.g., it can exhaustively list all possible latitudes of an aircraft within a certain flight range expressed in "degrees, minutes, and seconds," and its total number can be considered the size of the latitude vocabulary), D ll This indicates the dimension of the joint embedding of longitude and latitude. Furthermore, within the brackets of `Embedding`, `lon` represents the attribute value of the flight attribute "longitude" in the attribute information of the observed waypoint, and `lat` represents the attribute value of the flight attribute "latitude" in the attribute information of the observed waypoint. Therefore, this method can combine longitude and latitude into the same embedding vector space for embedding mapping, which helps to learn the joint relationship between longitude and latitude while learning longitude and latitude features. Similarly, the second embedding is represented by `coun_type`. emb It can be represented as: (5) In the above formula (5), Embedding represents (coun_size + type_size). D ct The second embedding matrix is ​​of size, where `country_size` represents the size of the country-specific vocabulary (e.g., all possible countries of the aircraft can be enumerated, and their total number can be considered the size of the country-specific vocabulary), and `type_size` represents the size of the aircraft type vocabulary (e.g., all possible aircraft types can be enumerated, and their total number can be considered the size of the aircraft type vocabulary), D ct This indicates the dimension of the joint embedding of country and aircraft type. Furthermore, the `coun` within parentheses of `Embedding` represents the attribute value of the flight attribute "country" in the attribute information of the observed waypoint, and `type` represents the attribute value of the flight attribute "aircraft type" in the attribute information of the observed waypoint. Therefore, this method can combine country and aircraft type into the same embedding vector space for embedding mapping, which helps to learn the joint relationship between country and aircraft type while learning country and aircraft type features. Similarly, the third embedding is represented by `sog_cog`. emb It can be represented as: (6) In the above formula (6), Embeding represents (sog_size + cog_size). D scThe third embedding matrix is ​​of size, where sog_size represents the size of the airspeed vocabulary (e.g., all possible airspeed values ​​of the aircraft can be enumerated, and their total number can be considered as the size of the airspeed vocabulary), cog_size represents the size of the heading vocabulary (e.g., all possible heading angle values ​​of the aircraft can be enumerated, and their total number can be considered as the size of the heading vocabulary), D sc This indicates the dimension of the joint embedding of heading and speed. Furthermore, within the brackets of `Embedding`, `sog` represents the attribute value of the flight attribute "speed" in the attribute information of the observed waypoint, and `cog` represents the attribute value of the flight attribute "heading" in the attribute information of the observed waypoint. Additionally, for the attribute value of the flight attribute "altitude" in the attribute information of the observed waypoint, it can be embedded using the fourth embedding matrix of the flight attribute "altitude" to obtain the fourth embedding representation. For details, please refer to the aforementioned embedding mappings of other flight attributes, which will not be repeated here. Therefore, the above method can combine heading and speed into the same embedding vector space for embedding mapping, which helps to learn the joint relationship between heading and speed while learning heading and speed features. Of course, the above example is only one possible example of joint embedding representation when several flight attributes, including longitude, latitude, altitude, heading, speed, country, and aircraft type, are used. Other possible joint embedding methods are not limited here, nor will they be listed one by one. The above method, by jointly embedding and representing several flight attributes, can not only learn the characteristics of the flight attributes themselves, but also model the relationships between flight attributes, which helps to fully mine the attribute values ​​of several flight attributes in the attribute information of observed track points.

[0017] Step S12: Perform semantic extraction based on the embedded feature representation to obtain the semantic feature representation.

[0018] Specifically, semantic extraction can be performed by a feature extraction network, which may include, but is not limited to, Transformer blocks. The network structure of the feature extraction network is not limited here. For example, the feature extraction network may include an L-layer (e.g., 8-layer) Transformer block. As a possible example, taking a feature extraction network including Transformer blocks, positional embedding representations can also be obtained (e.g., through positional encoding methods such as rotational positional encoding; for details, please refer to the technical details of positional encoding, which will not be elaborated here). The embedded feature representation and the positional embedding representation are then input together into the feature extraction network for semantic feature extraction to obtain the semantic feature representation (e.g., the output feature representation of the last hidden layer of the feature extraction network can be used as the semantic feature representation). Of course, the above examples are only a few possible examples of semantic extraction in practical applications; other possible implementation methods are not limited here, nor will they be listed one by one.

[0019] Step S13: For each flight attribute: perform route prediction based on semantic feature representation to obtain the route probability of each expert network in the expert network group applicable to the flight attribute, and obtain the target feature representation of the flight attribute based on the route probability of each expert network and the output feature representation processed by each expert network respectively based on the semantic feature representation, and perform mapping prediction based on the target feature representation to obtain the attribute value of the flight attribute in the predicted waypoint.

[0020] In this embodiment of the disclosure, the expert network group applicable to flight attributes may include: a shared expert network applicable to all flight attributes and a customized expert network specifically for flight attributes. It should be noted that the shared expert network can be used to learn general temporal features of flight trajectories, while the customized expert network specifically for flight attributes can be used to learn corresponding features of flight attributes. For example, when several flight attributes include longitude and latitude, the expert network group applicable to longitude and the expert network group applicable to latitude can share a customized expert network specifically for longitude and latitude (e.g., a spatial expert network), which can be specifically responsible for longitude and latitude and learning geographical constraints. As another example, when several flight attributes include country and aircraft type, the expert network group applicable to country and the expert network group applicable to aircraft type can share a customized expert network specifically for country and aircraft type (e.g., a country-aircraft-type expert network), which can be specifically responsible for modeling country and aircraft type. Yet another example, when several flight attributes include heading and speed, the expert network group applicable to heading and the expert network group applicable to speed can share a customized expert network specifically for heading and speed (e.g., a dynamics expert network), which can be specifically responsible for heading and speed and learning physical maneuvering logic. Finally, when several flight attributes include altitude, the expert network group applicable to altitude can include a customized expert network specifically for altitude (e.g., an altitude expert network), which can be specifically responsible for vertical changes and learning barometric pressure or climb limits. Of course, the above examples are just a few possible examples in practical applications. Other possible situations are not limited here (for example, the seven flight attributes mentioned above can each have their own customized expert networks), and will not be listed one by one.

[0021] In one implementation scenario, please refer to Figure 2 , Figure 2 This is a schematic diagram illustrating an embodiment of the trajectory prediction method of this application. Figure 2As shown, the attribute information of subsequent predicted waypoints can be obtained by the waypoint prediction model from the attribute information of the observed waypoints. The waypoint prediction model can include: a shared expert network applicable to all flight attributes and customized expert networks specifically for each flight attribute. In cases where several flight attributes still include the aforementioned seven flight attributes, taking the aforementioned customized expert network configuration as an example, the waypoint prediction model can include one set of shared expert networks and four sets of customized expert networks (i.e., one set of customized expert networks specifically for longitude and latitude, one set of customized expert networks specifically for country of origin and aircraft type, one set of customized expert networks specifically for heading and speed, and one set of customized expert networks specifically for altitude). Of course, the above configuration of the waypoint prediction model is only one possible example. Here, the configuration of the waypoint prediction model is not limited to cases where several flight attributes are different or where customized expert networks are configured differently, nor will it be listed in detail.

[0022] In one implementation scenario, the expert network group applicable to flight attributes can contain one or more shared expert networks (e.g., two, three, or more than three); similarly, the expert network group applicable to flight attributes can contain one or more customized expert networks (e.g., two, three, or more than three). For ease of description, the shared expert networks contained in the expert network group applicable to flight attributes can be denoted as n, and the customized expert networks can be m.

[0023] In one implementation scenario, please refer to [link / reference]. Figure 2 For any given flight attribute, route prediction can be performed by a route prediction network applicable to that flight attribute. In other words, the expert network group applicable to a flight attribute can also include a route prediction network for that flight attribute. Taking several flight attributes including the aforementioned seven flight attributes as an example, the expert network group applicable to the flight attribute "longitude" can also include a route prediction network for the flight attribute "longitude"; the expert network group applicable to the flight attribute "latitude" can also include a route prediction network for the flight attribute "latitude"; the expert network group applicable to the flight attribute "heading" can also include a route prediction network for the flight attribute "heading"; the expert network group applicable to the flight attribute "speed" can also include a route prediction network for the flight attribute "speed"; the expert network group applicable to the flight attribute "country" can also include a route prediction network for the flight attribute "country"; the expert network group applicable to the flight attribute "aircraft type" can also include a route prediction network for the flight attribute "aircraft type"; and the expert network group applicable to the flight attribute "altitude" can also include a route prediction network for the flight attribute "altitude". Of course, the above example is only one possible example of implementing route prediction in practical applications. Other possible implementation methods are not limited here, nor will they be listed one by one.

[0024] In a specific implementation scenario, a route prediction network applicable to any flight attribute can specifically include a fully connected layer and a softmax layer. The fully connected layer can be used to further process the semantic feature representation to obtain the logits output by the fully connected layer (containing the values ​​corresponding to the m+n expert networks applicable to this flight attribute, representing the probability of selecting these expert networks respectively). The softmax layer can be used to normalize the logits output by the fully connected layer to obtain the routing probability of each expert network in the expert network group applicable to the flight attribute.

[0025] In a specific implementation scenario, for ease of description, for any flight attribute, the routing probability p of the i-th expert network in the expert network group applicable to this flight attribute can be represented as... i (x) is represented as: (7) In the above formula (7), h(x) i h(x) j Let each represent the logits value (representing the probability of selecting each of the m+n expert networks applicable to this flight attribute) respectively. For example, it can be expressed as: (8) In the above formula (8), The network parameters represent the route prediction network applicable to this flight attribute; that is, it can be represented as a real matrix of dimension d and size m+n. Furthermore, x can be represented as the semantic feature representation of the observed waypoints.

[0026] In one implementation scenario, the target feature representation can be obtained by weighting and summing the output feature representations obtained by the expert network processing the semantic feature representation using the routing probabilities of the expert network. That is, for any flight attribute, the target feature representation for this flight attribute can be obtained by weighting the output feature representations obtained by processing the semantic feature representation through each expert network in the expert network group applicable to this flight attribute, based on the routing probabilities of each expert network in the expert network group applicable to this flight attribute. For example, the expert network may include fully connected layers (e.g., it may include two fully connected layers). In this case, for any expert network, the output feature representation E(x) of the semantic feature representation x processed by the expert network can be expressed as: (9) In the above formula (9), W1 and W2 represent the weight parameters of the first and second fully connected layers, respectively, b1 and b2 represent the bias parameters of the first and second fully connected layers, respectively, and σ represents the activation function. Based on this, for any flight attribute, the target feature representation of this flight attribute can be obtained by weighting the output feature representation of the semantic feature representation processed by each expert network in the expert network group applicable to this flight attribute, based on the routing probabilities of each expert network in the expert network group applicable to this flight attribute. For ease of description, the target feature representation y for any flight attribute can be expressed as: (10) In the above formula, p i (x) represents the routing probability of the i-th custom expert network, E i (x) represents the output feature representation of the i-th custom expert network, p j (x) represents the routing probability of the j-th shared expert network, E j (x) represents the output feature representation of the j-th shared expert network. Ψ represents the custom expert network, and Φ represents the shared expert network. As a possible example, for any flight attribute, the output feature representation obtained by processing the semantic feature representation through all expert networks in the expert network group applicable to this flight attribute can be weighted based on the routing probability of all expert networks in the expert network group applicable to this flight attribute to obtain the target feature representation of this flight attribute. That is, in the above formula (10), Ψ can specifically represent all custom expert networks in the expert network group applicable to this flight attribute, and Φ can specifically represent all shared expert networks in the expert network group applicable to this flight attribute. As another possible example, for any flight attribute, the target feature representation of this flight attribute can be obtained by weighting the output feature representation of the semantic feature representation processed by some expert networks in the expert network group applicable to this flight attribute, based on the routing probabilities of some expert networks in the expert network group applicable to this flight attribute (e.g., the top-k customized expert networks and the top-k shared expert networks can be selected according to the routing probabilities). That is, in the above formula (10), Ψ can specifically represent the top-k customized expert networks in the expert network group applicable to this flight attribute, and Φ can specifically represent the top-k shared expert networks in the expert network group applicable to this flight attribute. Of course, the above examples are only a few possible examples in practical applications. Other possible weighting methods are not limited here, nor will they be listed one by one.

[0027] In one implementation scenario, please refer to [link / reference]. Figure 2For any flight attribute, the expert network group applicable to this flight attribute may further include a mapping prediction network, which performs mapping prediction on the target feature representation of this flight attribute to obtain the attribute value of this flight attribute in the predicted waypoints at subsequent time steps. For example, the mapping prediction network may include, but is not limited to, a multilayer perceptron, etc., and the network structure of the mapping prediction network is not limited here.

[0028] Step S14: Based on the attribute values ​​of each flight attribute in the predicted waypoint, obtain the attribute information of the predicted waypoint.

[0029] In this embodiment of the disclosure, the predicted waypoint can be located at a time after the observed waypoint. Exemplarily, the predicted waypoint can be located at the next time after the observed waypoint. Exemplarily, this embodiment of the disclosure can be executed based on the attribute information of the observed waypoint at time t to obtain the attribute values ​​of each flight attribute in the predicted waypoint at time t+1. Then, the attribute values ​​of each flight attribute in the predicted waypoint at time t+1 can be combined to obtain the attribute information of the predicted waypoint at time t+1. As a possible example, after obtaining the attribute information of the predicted waypoint, the attribute information of the new observed waypoint at the current time can be further obtained, and the aforementioned step of feature embedding based on the attribute information of the observed waypoint to obtain the embedded feature representation can be returned to obtain the attribute information of the new predicted waypoint. For example, taking the aforementioned example again, when the attribute information of the predicted track point at time t+1 is obtained, time t+1 has not yet actually arrived, so as to guide or control the subsequent flight of the aircraft. When time t+1 arrives, the attribute information of the observed track point at time t+1 (i.e., the aforementioned new observed track point) can be obtained, and the above-described operation of this embodiment can be performed again to obtain the attribute information of the predicted track point at time t+2. This process is repeated, so that the aircraft can be continuously guided or controlled during its flight. Of course, the above example is only one possible example in actual application, and other possible situations are not limited here (e.g., the predicted track point may only be used for research purposes and not for actual guidance or control of the aircraft's flight process), and will not be listed in detail.

[0030] In one implementation scenario, as mentioned earlier, the attribute information of the predicted waypoints can be obtained by the waypoint prediction model from the attribute information of the observed waypoints. As a possible example, the model structure of the waypoint prediction model can be found in [reference needed]. Figure 2The aforementioned descriptions will not be repeated here. Based on this, in order to train the trajectory prediction model, during the prediction of the attribute information of each sample observation trajectory point in the training batch, for each flight attribute: the sample routing probability and the number of times each expert network in the expert network group applicable to the flight attribute can be obtained. Furthermore, based on the sample routing probability and the number of times the same expert network is selected in each prediction in the training batch, the preference degree of the flight attribute for the expert network can be obtained. The load balancing degree of the expert network group applicable to the flight attribute is obtained by fusing the preference degree of the flight attribute for each expert network in the expert network group applicable to the flight attribute. In the process of predicting the attribute information of any sample observation trajectory point, the expert network with the highest sample routing probability can be considered selected (e.g., in the process of predicting the attribute information of any sample observation trajectory point, for any flight attribute, the expert network with the highest sample routing probability in the expert network group applicable to this flight attribute can update its number of times it is selected once, i.e., the number of times this expert network is selected can be incremented by 1). Based on this, the load balancing loss can be obtained from the load balancing degree of expert network groups applicable to each flight attribute. Then, the network parameters of the trajectory prediction model can be adjusted based on this load balancing loss. This method, by measuring the load balancing loss during the training process of the trajectory prediction model, can prevent individual expert networks from becoming route preferences that prevent all expert networks from being fully trained and applied, thus helping to ensure load balancing among expert networks as much as possible.

[0031] In a specific implementation scenario, the specific meaning of the sample observation track points can be found in the aforementioned description of observation track points, and will not be repeated here. Furthermore, the specific process of predicting the attribute information of the sample observation track points using the track prediction model can be found in the aforementioned description of predicting observation track points, and will not be repeated here. The specific process of obtaining the sample routing probability and the number of times a sample is selected can also be found in the aforementioned description, and will not be repeated here. For ease of description, the training batch can be denoted as B, and the total number of sample observation track points in the training batch can be denoted as T.

[0032] In a specific implementation scenario, for any flight attribute, after obtaining the sample routing probabilities and the number of times each expert network in the expert network group applicable to this flight attribute is selected, the degree of preference of the flight attribute for the expert network can be measured. For example, a first preference can be obtained by averaging the sample routing probabilities predicted by the expert network in each training batch, and a second preference can be obtained by averaging the number of times each prediction is selected by the expert network in each training batch to the total number of samples in the training batch. Based on this, the load balancing degree of the flight attribute for the expert network can be obtained by multiplying the first and second preferences of the same expert network. For ease of description, taking flight attribute g as an example, the total number of expert networks in the expert network group applicable to flight attribute g can be denoted as N. Then, for the i-th expert network, the preference can be calculated based on the sample routing probabilities p predicted in each training batch B. i (x) is averaged to obtain the first preference level: (11) In the above formula (11), x represents the sample observation track point, T represents the total number of sample observation track points in the sample batch, and p i (x) represents the sample routing probability of the i-th expert network in the expert network group applicable to flight attribute g when predicting the sample observation track point x. Furthermore, for the i-th expert network, the second preference can be obtained based on the ratio of the number of times this expert network's predictions are selected in each training batch to the total number of samples in the training batch: (12) In the above formula (12), argmaxp(x)=i indicates that the sample routing probability of the i-th expert network is the highest, and I{argmaxp(x)=i} represents the statistical number of times the i-th expert network is selected. Based on this, the first preference degree and the second preference degree of the same expert network can be multiplied to obtain the preference degree of the corresponding expert network. Then, the preference degrees of each expert network in the expert network group applicable to the flight attribute are accumulated to obtain the load balancing degree of the flight attribute for the expert network. In this way, the load balancing loss can be obtained based on the load balancing degree of the expert network group applicable to each flight attribute (e.g., summed). For ease of description, the load balancing degree of the flight attribute g for the expert network can be expressed as: (13) In the above formula (13), Let represent the first preference and second preference of flight attribute g with respect to the i-th expert network, respectively. For details, please refer to the aforementioned descriptions, which will not be repeated here. Based on this, for all flight attributes, the load balancing loss L can be obtained by summing the aforementioned load balancing degrees. load_balance : (14) In the above formula (14), α represents the multiplication coefficient, which is used to control the proportion of load balancing loss in the total loss function, so as to ensure that it can be load balanced as much as possible and not affect the role of the model's main loss (see the following related description). The above method is based on averaging the sample route probabilities predicted by the expert network in each training batch to obtain the first preference degree, and based on the ratio of the statistical number of times the expert network's predictions were selected in each training batch to the total number of samples in the training batch to obtain the second preference degree. Then, the first preference degree and the second preference degree of the same expert network are multiplied to obtain the degree of preference of the flight attribute for the expert network. This can promote the uniform distribution of various flight attributes of the sample observation track points among different expert networks as much as possible.

[0033] In a specific implementation scenario, the observed track points can be labeled with the attribute information of a first sample track point. This first sample track point can be located at a later time (e.g., a subsequent time). Before adjusting the network parameters of the track prediction model based on the load balancing loss, the attribute information of the observed track points can be predicted using the track prediction model to obtain the attribute information of a second sample track point. This second sample track point can also be located at a later time (e.g., a subsequent time). It should be noted that during training, the track prediction model can be constrained to predict the time following the input data for each prediction. Based on this, the main loss of the model can be obtained based on the difference between the attribute information of the first and second sample track points. Then, the network parameters of the track prediction model can be adjusted based on the main loss and the load balancing loss. For example, the cross-entropy loss function can be used to measure the difference between the attribute information of the first sample waypoint and the attribute information of the second sample waypoint, thus obtaining the model's main loss (e.g., the difference between the attribute values ​​of the same flight attribute in the attribute information of the first and second sample waypoints can be measured first to obtain the sub-loss of the corresponding flight attribute, and then the sub-losses of each flight attribute can be weighted, summed, and fused to obtain the model's main loss). For ease of description, the model's main loss can be expressed as: (15) In the above formula (15), CE represents the cross-entropy loss function, h T+1 P represents the attribute information of the first sample track point. T+1This can represent the attribute information of the second sample waypoint. Furthermore, M represents the total number of flight attributes, and w... i L represents the loss weight of the i-th flight attribute. i Let represent the sub-loss of the i-th flight attribute. The above method predicts the attribute information of the sample observed track points based on the track prediction model, obtaining the attribute information of the second sample track point. Based on the difference between the attribute information of the first and second sample track points, the main loss of the model is obtained. Then, based on the main loss and the load balancing loss, the network parameters of the track prediction model are adjusted. This allows the main loss to ensure the basic performance of the track prediction model as much as possible, while the load balancing loss further promotes the even distribution of various flight attributes of the sample observed track points among different expert networks. It should be noted that the above implementation is only one possible training method for the track prediction model in practical applications. Other possible training methods are not limited here (e.g., the load balancing loss can be ignored, i.e., training is based on the main loss), and will not be listed in detail. Furthermore, when the aforementioned embedded feature representation is obtained by embedding mapping using the aforementioned embedding matrix, the adjustment of the network parameters of the track prediction model based on the main loss and the load balancing loss can also be based on the main loss and the load balancing loss. The aforementioned embedding matrices should be adjusted accordingly. For details, please refer to the technical details related to the optimizer for model training, which will not be elaborated here.

[0034] In one implementation scenario, as mentioned earlier, the attribute information of the predicted waypoints can be obtained by the waypoint prediction model predicting the attribute information of the observed waypoints. The waypoint prediction model can be trained based on the model's main loss, which is obtained by weighting the sub-losses of each flight attribute involved in the attribute information and their respective loss weights. It should be noted that the loss weights of the flight attributes can be negatively adjusted based on the optimization speed of the corresponding sub-losses. That is, for any flight attribute, if the optimization speed of the sub-losses of the flight attribute is fast, its loss weight can be relatively small; conversely, if the optimization speed of the sub-losses of the flight attribute is slow, its loss weight can be relatively large, so that each flight attribute can be trained and optimized as fully as possible. To facilitate understanding of the negative adjustment of loss weights based on the optimization speed of sub-losses, a fixed loss weight is used as a reference. If the loss weights remain constant, during training, updates of flight attributes with large gradients will dominate the model, resulting in fast training but potentially overfitting (e.g., if the vocabulary size for longitude and latitude is much larger than that for country and aircraft type, the gradient norm will differ significantly when calculating the corresponding main loss after feature embedding). Conversely, for flight attributes with small gradients, the learning speed is slow, and performance may degrade. Therefore, negatively adjusting the loss weights based on the optimization speed of sub-losses ensures that each flight attribute has similar training intensity for the gradients of shared parameters. As a possible example, to negatively adjust the loss weights, various flight attributes can be selected as the current attributes (e.g., the aforementioned seven flight attributes "longitude," "latitude," "speed," "heading," "country," "aircraft type," and "altitude" can be selected as the current attributes). Based on the loss weights of the flight attributes adjusted in the previous training round and the sub-losses from the previous training round, the actual gradient norm of the flight attributes is obtained. Then, based on the ratio of the sub-losses of the flight attributes in this training round to those in the first training round, the optimization speed of the flight attributes is obtained. Thus, based on the baseline gradient norm obtained by fusing the actual gradient norms of each flight attribute and the optimization speed of the current attribute, the expected gradient norm of the current attribute can be obtained. Finally, based on the difference between the expected gradient norm and the actual gradient norm of the current attribute, the loss weights of the current attribute adjusted in the previous training round are adjusted, resulting in the adjusted loss weights of the current attribute in this training round. This method, by measuring the actual gradient norm and the optimization speed to obtain the expected gradient norm, and then adjusting the loss weights based on the difference between the expected and actual gradient norms, can ensure that each flight attribute receives balanced training as much as possible.

[0035] In a specific implementation scenario, for the i-th flight attribute, its actual gradient norm G i It can be represented as: (16) In the above formula (16), For gradient operators, it means that the gradient is shared with respect to the underlying parameters W. s (For example, the network parameters of the last layer of the backbone network in a trajectory prediction network) Calculate the partial derivatives. Furthermore, w i L represents the loss weight of the i-th flight attribute after adjustment in the previous training round. i This represents the sub-loss of the i-th flight attribute in the previous training round.

[0036] In a specific implementation scenario, as one possible example, to obtain the optimization speed of the flight attribute, one can directly obtain the ratio of the sub-loss of the flight attribute in the current training round to that in the first training round, as the optimization speed of the flight attribute. Alternatively, as another possible example, to obtain the optimization speed of the flight attribute, one can first obtain the ratio of the sub-loss of the flight attribute in the current training round to that in the first training round, as the loss ratio, and then compare these loss ratios for each flight attribute to obtain the optimization speed of each flight attribute. For ease of description, we will still take the i-th flight attribute as an example, whose optimization speed r... i It can be represented as: (17) In the above formula (17), L i (t), L j (t) represent the sub-losses of the i-th and j-th flight attributes respectively in this round of training (i.e., the t-th round of training), L i (0), L j (0) represents the sub-loss of the i-th flight attribute and the j-th flight attribute in the first round of training, respectively. Furthermore, M represents the total number of flight attributes. When calculating the optimization speed using the above formula (17), if the optimization speed r of the i-th flight attribute... i If the value is less than 1, it indicates that the i-th flight attribute is learned relatively quickly; conversely, if the optimization speed r of the i-th flight attribute is less than 1, it indicates that the learning speed is relatively fast. i If the value is greater than 1, it indicates that the learning speed of the i-th flight attribute is relatively slow. Of course, if the optimization speed r of the i-th flight attribute is... i =1 indicates that the learning of the i-th flight attribute is moderate (i.e., at the average level).

[0037] In a specific implementation scenario, after obtaining the actual gradient norms of each flight attribute, the actual gradient norms of each flight attribute can be averaged to obtain the baseline gradient norm. For ease of description, the baseline gradient norm is... It can be represented as: (18) In the above formula, G iLet represent the actual gradient norm of the i-th flight attribute, and M represent the total number of flight attributes. Based on this, the expected gradient norm of the current attribute can be obtained by scaling the baseline gradient norm using a first scaling factor with the optimization speed of the current attribute as the base and the balancing hyperparameter as the exponent. For ease of description, let's take the i-th flight attribute as the current attribute as an example. The expected gradient norm of the current attribute... It can be represented as: (19) In the above formula (19), β represents the balancing hyperparameter, used to control the penalty for imbalances in learning rate. The larger β is, the stronger the balancing force. Therefore, if a certain flight attribute learns slowly (r... i If the gradient norm is greater than 1, then it would be desirable to increase the gradient norm, and if a certain flight attribute is learned quickly (r... i If <1, then it is expected that the gradient norm will be reduced.

[0038] In a specific implementation scenario, after obtaining the expected gradient norm, the loss weights of the current attribute, adjusted in the previous training iteration, can be adjusted based on the difference between the expected and actual gradient norms of the current attribute, resulting in the adjusted loss weights for the current attribute in this training iteration. For example, the first regularization loss of the current attribute can be obtained based on the difference between the expected and actual gradient norms. Then, the product of the partial derivative of the first regularization loss with respect to the loss weights and the learning rate is used as the adjustment range of the loss weights for the current attribute in this training iteration. Furthermore, based on this adjustment range, the loss weights of the current attribute, adjusted in the previous training iteration, can be adjusted to obtain the adjusted loss weights for the current attribute in this training iteration (when each flight attribute is selected as the current attribute, the adjusted loss weights for each flight attribute in this training iteration can be obtained). These adjusted loss weights are used to weight the sub-losses in the next training iteration to obtain the main loss of the model in the next training iteration. For ease of description, taking the i-th flight attribute as the current attribute as an example, the first regularization loss can be expressed as: (20) Based on this, we can obtain the first regularization loss of the i-th flight attribute and its weight w. i The product of the partial derivative of the i-th flight attribute and the learning rate η is used as the adjustment range of the loss weight of the i-th flight attribute in this round of training. Based on this, the loss weight of the i-th flight attribute after the adjustment in the previous round of training is adjusted to obtain the loss weight of the i-th flight attribute after the adjustment in this round of training: (twenty one) In the above formula (21), the arrow points to w iThis represents the loss weight of the i-th flight attribute after this round of training, where w is not pointed to by the arrow. i This represents the loss weight of the i-th flight attribute after adjustment in the previous training round. It should be noted that η is the learning rate for updating the loss weights, which can be set differently from the learning rate used to train the model parameters. Of course, if needed, it can also be set to the same learning rate as the model parameters. After obtaining the loss weights of each flight attribute after adjustment in this training round through the above operations, a normalization operation can be performed on the loss weights of each flight attribute after adjustment in this training round.

[0039] In another implementation scenario, unlike the aforementioned method of adjusting loss weights, after obtaining the expected gradient norm of the current attribute based on the fusion of the actual gradient norms of each flight attribute and the optimization speed of the current attribute, it is not necessary to adjust the loss weights of the current attribute after the previous training round based on the difference between the expected gradient norm and the actual gradient norm of the current attribute. Instead, it is possible to first obtain the second regularization loss based on the difference between the expected gradient norm and the actual gradient norm of each flight attribute, and obtain the uncertainty loss based on the sub-loss of each flight attribute and the noise parameters updated after the previous training round. Then, the parameters are adjusted based on the second regularization loss, the uncertainty loss, and the main loss of the model to obtain the network parameters of the trajectory prediction model updated after this round of training and the noise parameters of each flight attribute updated after this round of training. Based on this, the rate of change of the current attribute with respect to the noise parameters obtained through the current training iteration can be obtained. This rate of change allows for updating the expected gradient norm of the current attribute, resulting in a new expected gradient norm that is negatively correlated with the rate of change. Furthermore, the difference between the new expected gradient norm and the actual gradient norm allows for adjusting the loss weights of the current attribute after the previous training iteration, resulting in the adjusted loss weights for the current attribute after this training iteration. This method, which uses the rate of change of the noise parameters to update the expected gradient norm to obtain a new expected gradient norm that is negatively correlated with the rate of change, and then adjusts the loss weights based on the difference between the new and actual expected gradient norms, fully considers the differences in noise distribution among different flight attributes, integrates uncertainty awareness with gradient balancing, and helps improve the balance of learning speed across various flight attributes.

[0040] In a specific implementation scenario, the method for obtaining the second regularization loss can be found in the aforementioned description of the first regularization loss, and will not be repeated here. Furthermore, to obtain the uncertainty loss, the sub-losses of the flight attribute are weighted based on the noise parameters obtained from the previous training iteration to obtain the first sub-loss of the flight attribute. The second sub-loss of the flight attribute is then obtained based on the noise parameters obtained from the previous training iteration, with the first sub-loss negatively correlated to the noise parameters and the second sub-loss positively correlated. Based on this, the first and second sub-losses of each flight attribute can be fused to obtain the uncertainty loss. For ease of description, the uncertainty loss L... uncertainty It can be represented as: (twenty two) In the above formula (22), σ i L represents the noise parameter obtained by updating the i-th flight attribute after the previous training round. i This represents the sub-loss of the i-th flight attribute. In the above formula (22), the term before the plus sign represents the first sub-loss, and the term after the plus sign represents the second sub-loss.

[0041] In a specific implementation scenario, after obtaining the uncertainty loss, the total loss can be obtained by weighting the second regularization loss, the uncertainty loss, and the model's main loss; alternatively, after obtaining the uncertainty loss, the total loss can also be obtained by weighting the second regularization loss, the uncertainty loss, the load balancing loss (see the aforementioned description), and the model's main loss. Based on this, parameters can be adjusted according to the total loss to obtain the network parameters updated in this round of training for the trajectory prediction model and the noise parameters updated in this round for each flight attribute. In other words, the noise parameters can be optimized and adjusted along with the network parameters. As another possible example, when the embedded feature representation is embedded and mapped using the aforementioned embedding matrix, the noise parameters and the embedding matrix can be optimized and adjusted along with the network parameters.

[0042] In a specific implementation scenario, after obtaining the noise parameters of the current attribute updated through this round of training, the rate of change of the current attribute with respect to the noise parameters can be obtained. Then, the expected gradient norm of the current attribute is updated based on this rate of change, resulting in a new expected gradient norm for the current attribute. For example, a second scaling factor for the current attribute can be obtained based on its rate of change with respect to the noise parameters, and this second scaling factor is negatively correlated with the rate of change. The expected gradient norm of the current attribute is then scaled based on this second scaling factor to obtain a new expected gradient norm for the current attribute. For ease of description, we will still take the i-th flight attribute as the current attribute as an example, and the new expected gradient norm of the current attribute... It can be represented as: (twenty three) In the above formula (23), dσ i / dt represents the rate of change of the current attribute with respect to the noise parameter; the arrow does not point to... The arrow points to the expected gradient norm of the current attribute. Let β represent the new expected gradient norm of the current attribute, and let β represent the balance factor. It can be seen that for any flight attribute, if its data itself is very noisy (high uncertainty), then even if it is currently learning slowly (small gradient), it should not be blindly accelerated to fit it, otherwise it is easy to overfit the noise. In this way, when the uncertainty of a certain flight attribute (i.e. the aforementioned noise parameter σ) drops sharply (i.e., it can be regarded as the model starting to overconfidently fit the noise), even if its gradient is small, its learning rate should be suppressed to prevent overfitting as much as possible. It should be noted that after obtaining the new expected gradient norm of the current attribute, the aforementioned implementation method (such as formulas (20)~(21) and their related textual descriptions) can be used to obtain the loss weight of the current attribute after this round of training (in the case of selecting each flight attribute as the current attribute, the loss weight of each flight attribute after this round of training can be obtained), which is used to weight its sub-loss in the next round of training to obtain the model main loss of the next round of training. For details, please refer to the aforementioned related descriptions, which will not be repeated here.

[0043] The above scheme embeds features based on the attribute information of observed waypoints to obtain embedded feature representations. The attribute information includes the attribute values ​​of several flight attributes. Semantic extraction is then performed based on the embedded feature representations to obtain semantic feature representations. For each flight attribute: route prediction is performed based on the semantic feature representations to obtain the routing probabilities of each expert network in the expert network group applicable to the flight attribute. Based on the routing probabilities of each expert network and the output feature representations processed by each expert network, the target feature representation of the flight attribute is obtained. Mapping prediction is then performed based on the target feature representations to obtain the attribute values ​​of the flight attributes in the predicted waypoints. The expert network group applicable to the flight attributes includes a shared expert network applicable to all flight attributes and a customized expert network specifically for flight attributes. Finally, based on the attribute values ​​of each flight attribute in the predicted waypoints, the attribute information of the predicted waypoints is obtained. Since the predicted waypoints are located after the observed waypoints, the path prediction... Instead of modeling it as a multi-class classification problem for each flight attribute, this approach predicts each flight attribute separately. Firstly, it performs route prediction based on semantic feature representations for each flight attribute, which helps to fully consider the unique characteristics of different flight attributes. Secondly, when predicting the attribute values ​​of each flight attribute, it relies on an expert network group containing a shared expert network and a custom expert network. This allows the custom expert network to fully learn the unique characteristics of different flight attributes, and the shared expert network to further learn the general temporal features of flight trajectories. Finally, using the routing probabilities of each expert network in the expert network group applicable to the flight attribute obtained during the route prediction process, as well as the output feature representations processed by each expert network, the target feature representation of the flight attribute is obtained. This approach avoids the predicament of different flight attributes being forced to share the same feature representation for attribute prediction, thus improving the accuracy of mapping and predicting the attribute values ​​of flight attributes. Therefore, it can improve the accuracy of trajectory prediction.

[0044] Please see Figure 3 , Figure 3This is a schematic diagram of the framework of an embodiment of the trajectory prediction device of this application. The trajectory prediction device 30 includes: a feature embedding module 31, a semantic extraction module 32, an attribute prediction module 33, and a trajectory acquisition module 34. The feature embedding module 31 is used to embed features based on the attribute information of observed trajectory points to obtain an embedded feature representation; wherein, the attribute information includes attribute values ​​of several flight attributes. The semantic extraction module 32 is used to extract semantics based on the embedded feature representation to obtain a semantic feature representation. The attribute prediction module 33 is used to perform route prediction based on the semantic feature representation for each flight attribute to obtain the routing probability of each expert network in the expert network group applicable to the flight attribute. Based on the routing probabilities and semantic feature representations of each expert network, and the output feature representations processed by each expert network, the target feature representation of the flight attribute is obtained. Based on the target feature representation, a mapping prediction is performed to obtain the attribute value of the flight attribute in the predicted track point. The expert network group applicable to the flight attribute includes: a shared expert network applicable to all flight attributes and a customized expert network specifically for flight attributes. The track acquisition module 34 is used to obtain the attribute information of the predicted track point based on the attribute value of each flight attribute in the predicted track point. The predicted track point is located after the observed track point.

[0045] In the above scheme, the trajectory prediction device 30 performs feature embedding based on the attribute information of the observed trajectory points to obtain embedded feature representations. The attribute information includes attribute values ​​of several flight attributes. Then, semantic extraction is performed based on the embedded feature representations to obtain semantic feature representations. For each flight attribute: route prediction is performed based on the semantic feature representations to obtain the routing probabilities of each expert network in the expert network group applicable to the flight attribute. Based on the routing probabilities of each expert network and the output feature representations processed by each expert network, the target feature representation of the flight attribute is obtained. Mapping prediction is performed based on the target feature representations to obtain the attribute values ​​of the flight attributes in the predicted trajectory points. The expert network group applicable to the flight attributes includes: a shared expert network applicable to all flight attributes and a customized expert network specifically for flight attributes. Then, based on the attribute values ​​of each flight attribute in the predicted trajectory points, the attribute information of the predicted trajectory points is obtained. The predicted trajectory points are located after the observed trajectory points. In trajectory prediction, each flight attribute is predicted separately, rather than being modeled as a multi-class classification problem for each flight attribute. Firstly, routing prediction is performed on each flight attribute based on semantic feature representations, which helps to fully consider the unique characteristics of different flight attributes. Secondly, when predicting the attribute values ​​of each flight attribute, a group of expert networks, including a shared expert network and a custom expert network, is relied upon. This allows the custom expert network to fully learn the unique characteristics of different flight attributes, and the shared expert network to further learn the general temporal features of the flight trajectory. Finally, using the routing probabilities of each expert network in the expert network group applicable to the flight attribute obtained during the routing prediction process, as well as the output feature representations processed by each expert network, the target feature representation of the flight attribute is obtained. This avoids the predicament of different flight attributes being forced to share the same feature representation for attribute prediction, thus improving the accuracy of mapping and predicting the attribute values ​​of the flight attributes. Therefore, the accuracy of trajectory prediction can be improved.

[0046] In some disclosed embodiments, the attribute information of the predicted track points is obtained by the track prediction model predicting the attribute information of the observed track points. The track prediction device 30 includes a load measurement module, which is used to: obtain the sample routing probability and the number of times each expert network in the expert network group applicable to the flight attribute is selected, and obtain the preference degree of the flight attribute for the expert network based on the sample routing probability and the number of times the same expert network is selected in each prediction in the training batch, and fuse the preference degree of the flight attribute for each expert network in the expert network group applicable to the flight attribute to obtain the load balance degree of the expert network group applicable to the flight attribute; wherein, in the process of predicting the attribute information of any sample observed track point, the expert network with the highest sample routing probability is selected; the track prediction device 30 includes a first calculation module, which is used to obtain the load balance loss based on the load balance degree of the expert network group applicable to each flight attribute; the track prediction device 30 includes a parameter adjustment module, which is used to adjust the network parameters of the track prediction model based on the load balance loss.

[0047] In some disclosed embodiments, the sample observation track points are labeled with attribute information of a first sample track point, which is located at a time after the sample observation track point. The track prediction device 30 includes a sample prediction module, which is used to predict the attribute information of the sample observation track points based on the track prediction model to obtain the attribute information of a second sample track point; wherein, the second sample track point is located at a time after the sample observation track point. The track prediction device 30 includes a second calculation module, which is used to obtain the model principal loss based on the difference between the attribute information of the first sample track point and the attribute information of the second sample track point. The parameter adjustment module is specifically used to adjust the network parameters of the track prediction model based on the model principal loss and the load balancing loss.

[0048] In some disclosed embodiments, the load metric module includes a preference metric submodule, which is used to average the sample route probabilities predicted by the expert network in each training batch to obtain a first preference degree, and to obtain a second preference degree based on the ratio of the statistical number of times the expert network's predictions were selected in each training batch to the total number of samples in the training batch; the load metric module includes a preference multiplication submodule, which is used to multiply the first preference degree and the second preference degree of the same expert network to obtain the degree of preference of the flight attribute for the expert network.

[0049] In some publicly disclosed embodiments, the attribute information of the predicted waypoints is obtained by the waypoint prediction model predicting the attribute information of the observed waypoints. The waypoint prediction model is trained based on the model's main loss. The model's main loss is obtained by weighting the sub-losses of each flight attribute involved in the attribute information and their loss weights. Furthermore, the loss weights of the flight attributes are negatively adjusted based on the optimization speed of the corresponding sub-losses.

[0050] In some disclosed embodiments, the trajectory prediction device 30 includes an attribute selection module for selecting various flight attributes as current attributes; the trajectory prediction device 30 includes a norm velocity module for obtaining the actual gradient norm of the flight attribute based on the loss weights adjusted after the previous training and the sub-losses in the previous training, and obtaining the optimization velocity of the flight attribute based on the ratio of the sub-losses of the flight attribute in this training to that in the first training; the trajectory prediction device 30 includes an expected norm module for obtaining the expected gradient norm of the current attribute based on the baseline gradient norm obtained by fusing the actual gradient norms of various flight attributes and the optimization velocity of the current attribute; the trajectory prediction device 30 includes a weight adjustment module for adjusting the loss weights of the current attribute after the previous training based on the difference between the expected gradient norm and the actual gradient norm of the current attribute, to obtain the loss weights of the current attribute after the current training.

[0051] In some disclosed embodiments, the expected norm module includes a baseline determination submodule for averaging the actual gradient norms of each flight attribute to obtain a baseline gradient norm; the expected norm module includes a norm scaling submodule for scaling the baseline gradient norm based on a first scaling factor with the optimization speed of the current attribute as the base and the balance hyperparameter as the exponent to obtain the expected gradient norm of the current attribute.

[0052] In some disclosed embodiments, the weight adjustment module includes a first regularization submodule, used to obtain the first regularization loss of the current attribute based on the difference between the expected gradient norm and the actual gradient norm of the current attribute; the weight adjustment module includes an amplitude calculation submodule, used to obtain the product of the partial derivative of the first regularization loss with respect to the loss weight and the learning rate, as the adjustment amplitude of the loss weight of the current attribute in this round of training; the weight adjustment module includes a weight adjustment submodule, used to adjust the loss weight of the current attribute after the previous round of training based on the adjustment amplitude of the loss weight of the current attribute in this round of training, to obtain the loss weight of the current attribute after the current round of training.

[0053] In some disclosed embodiments, the weight adjustment module includes a second regularization submodule for obtaining a second regularization loss based on the difference between the expected gradient norm and the actual gradient norm of each flight attribute; the weight adjustment module includes an uncertainty submodule for obtaining an uncertainty loss based on the sub-loss of each flight attribute and the noise parameters updated in the previous training iteration; the weight adjustment module includes a parameter adjustment submodule for adjusting parameters based on the second regularization loss, the uncertainty loss, and the model's main loss to obtain the network parameters of the trajectory prediction model updated in the current training iteration and the noise parameters of each flight attribute updated in the current training iteration; the weight adjustment module... The module includes a change calculation submodule, used to obtain the rate of change of the current attribute with respect to the noise parameters updated in the current training round; the weight adjustment module includes a norm update submodule, used to update the expected gradient norm of the current attribute based on the rate of change of the current attribute with respect to the noise parameters, to obtain a new expected gradient norm of the current attribute; wherein, the new expected gradient norm is negatively correlated with the rate of change; the weight adjustment submodule is specifically used to adjust the loss weight of the current attribute after the previous training round based on the difference between the new expected gradient norm and the actual gradient norm of the current attribute, to obtain the loss weight of the current attribute after the current training round.

[0054] In some disclosed embodiments, the uncertainty submodule includes a sub-loss calculation unit, used to weight the sub-loss of the flight attribute based on the noise parameters obtained from the previous training update of the flight attribute to obtain a first sub-loss of the flight attribute, and to obtain a second sub-loss of the flight attribute based on the noise parameters obtained from the previous training update of the flight attribute; wherein, the first sub-loss of the flight attribute is negatively correlated with the noise parameters obtained from the previous training update of the flight attribute, and the second sub-loss of the flight attribute is positively correlated with the noise parameters obtained from the previous training update of the flight attribute; the uncertainty submodule includes a sub-loss fusion unit, used to fuse the first sub-loss and the second sub-loss of each flight attribute to obtain the uncertainty loss.

[0055] In some disclosed embodiments, the norm update submodule includes a scaling calculation unit for obtaining a second scaling factor for the current attribute based on the rate of change of the current attribute with respect to the noise parameter; wherein the second scaling factor is negatively correlated with the rate of change; the norm update submodule includes a norm scaling unit for scaling the expected gradient norm of the current attribute based on the second scaling factor of the current attribute to obtain a new expected gradient norm of the current attribute.

[0056] In some disclosed embodiments, several flight attributes are divided into several attribute combinations according to whether different flight attributes are related. The attribute values ​​of each flight attribute in the same attribute combination are embedded and mapped by the same embedding matrix to obtain the embedding representation of the corresponding attribute combination. The embedding feature representation is obtained by fusing the embedding representations of several attribute combinations.

[0057] In some disclosed embodiments, the feature embedding module 31 includes an embedding mapping submodule, used to perform embedding mapping on the attribute values ​​of longitude and latitude in the attribute information of the observed track point based on a first embedding matrix to obtain a first embedding representation; and to perform embedding mapping on the attribute values ​​of China and aircraft type in the attribute information of the observed track point based on a second embedding matrix to obtain a second embedding representation; to perform embedding mapping on the attribute values ​​of heading and speed in the attribute information of the observed track point based on a third embedding matrix to obtain a third embedding representation; and to perform embedding mapping on the attribute value of altitude in the attribute information of the observed track point based on a fourth embedding matrix to obtain a fourth embedding representation; feature embedding The input module 31 includes an embedding fusion submodule, which is used to fuse the first embedding representation, the second embedding representation, the third embedding representation and the fourth embedding representation to obtain the embedded feature representation; wherein, the number of rows of the first embedding matrix is ​​the sum of the sizes of the longitude term list and the latitude term list, and the number of columns of the first embedding matrix represents the dimension of the joint embedding of longitude and latitude; the number of rows of the second embedding matrix is ​​the sum of the sizes of the country term list and the aircraft type term list, and the number of columns of the second embedding matrix represents the dimension of the joint embedding of country and aircraft type; the number of rows of the third embedding matrix is ​​the sum of the sizes of the airspeed term list and the heading term list, and the number of columns of the third embedding matrix represents the dimension of the joint embedding of airspeed and heading.

[0058] In some disclosed embodiments, the trajectory prediction device 30 includes a new trajectory acquisition module for acquiring attribute information of new observed trajectory points at the current time; the trajectory prediction device 30 includes a loop iteration module for returning the step of embedding features based on the attribute information of observed trajectory points to obtain embedded feature representations, so as to obtain attribute information of new predicted trajectory points.

[0059] In some disclosed embodiments, the observed waypoints and predicted waypoints belong to the same aircraft; and / or, several flight attributes include at least one of the following: longitude, latitude, altitude, heading, speed, country of origin, and aircraft type; and / or, when several flight attributes include longitude and latitude, the expert network group applicable to longitude and the expert network group applicable to latitude share a customized expert network specifically for longitude and latitude; and / or, when several flight attributes include heading and speed, the expert network group applicable to heading and the expert network group applicable to speed share a customized expert network specifically for heading and speed; and / or, when several flight attributes include country of origin and aircraft type, the expert network group applicable to country of origin and the expert network group applicable to aircraft type share a customized expert network specifically for country of origin and aircraft type; and / or, in several When flight attributes include altitude, the expert network group applicable to altitude includes a custom expert network specifically for altitude; and / or, the attribute information of the predicted waypoints is obtained by the waypoint prediction model predicting the attribute information of the observed waypoints, the waypoint prediction model including: a shared expert network applicable to all flight attributes and custom expert networks specifically for each flight attribute; and / or, semantic extraction is performed by a feature extraction network, the feature extraction network including Transformer blocks; and / or, for any flight attribute, route prediction is performed by a route prediction network applicable to the corresponding flight attribute; and / or, the expert network includes fully connected layers; and / or, the target feature representation is obtained by weighted summation of the route probabilities of the expert network and the output feature representation obtained by the expert network processing the semantic feature representation.

[0060] Please see Figure 4 , Figure 4 This is a schematic diagram of a framework of an embodiment of the electronic device of this application. The electronic device 40 includes at least a memory 41 and a processor 42 coupled to each other. The memory 41 stores at least program instructions, and the processor 42 is used to execute the program instructions to implement the steps in any of the above-described trajectory prediction method embodiments. For details, please refer to the foregoing disclosed embodiments, which will not be repeated here. As one possible example, the electronic device 40 can be a master control device; or, as another possible example, the electronic device 40 can also be an aircraft such as a drone. In this case, the electronic device 40 can also include, but is not limited to, motors (not shown), rotors (not shown), etc. The specific type of the electronic device 40 is not limited here.

[0061] Specifically, processor 42 controls itself and memory 41 to implement the steps in any of the above-described trajectory prediction method embodiments. Processor 42 can also be referred to as a CPU (Central Processing Unit). Processor 42 may be an integrated circuit chip with signal processing capabilities. Processor 42 can also be a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor. Furthermore, processor 42 can be implemented using integrated circuit chips.

[0062] In the above scheme, the electronic device 40 performs feature embedding based on the attribute information of the observed flight path points to obtain embedded feature representations. The attribute information includes the attribute values ​​of several flight attributes. Then, semantic extraction is performed based on the embedded feature representations to obtain semantic feature representations. For each flight attribute: route prediction is performed based on the semantic feature representations to obtain the routing probabilities of each expert network in the expert network group applicable to the flight attribute. Based on the routing probabilities of each expert network and the output feature representations processed by each expert network, the target feature representation of the flight attribute is obtained. Mapping prediction is performed based on the target feature representations to obtain the attribute values ​​of the flight attributes in the predicted flight path points. The expert network group applicable to the flight attributes includes: a shared expert network applicable to all flight attributes and a customized expert network specifically for flight attributes. Then, based on the attribute values ​​of each flight attribute in the predicted flight path points, the attribute information of the predicted flight path points is obtained. The predicted flight path points are located after the observed flight path points. In trajectory prediction, each flight attribute is predicted separately, rather than modeling it as a multi-class classification problem for each flight attribute. Firstly, routing prediction is performed on each flight attribute based on semantic feature representations, which helps to fully consider the unique characteristics of different flight attributes. Secondly, when predicting the attribute values ​​of each flight attribute, it relies on an expert network group including a shared expert network and a custom expert network. This allows the custom expert network to fully learn the unique characteristics of different flight attributes, and the shared expert network to further learn the general temporal features of flight trajectories. Finally, using the routing probabilities of each expert network in the expert network group applicable to the flight attribute obtained during the routing prediction process, as well as the output feature representations processed by each expert network, the target feature representation of the flight attribute is obtained. This avoids the predicament of different flight attributes being forced to share the same feature representation for attribute prediction, thus improving the accuracy of mapping and predicting the attribute values ​​of flight attributes. Therefore, it can improve the accuracy of trajectory prediction.

[0063] Please see Figure 5 , Figure 5 This is a schematic diagram of a framework of an embodiment of the computer-readable storage medium of this application. The computer-readable storage medium 50 stores program instructions 51 that can be executed by a processor. The program instructions 51 are used to implement the steps in any of the above-described embodiments of the trajectory prediction method.

[0064] In the above scheme, the computer-readable storage medium 50 performs feature embedding based on the attribute information of the observed flight path points to obtain embedded feature representations. The attribute information includes attribute values ​​of several flight attributes. Then, semantic extraction is performed based on the embedded feature representations to obtain semantic feature representations. For each flight attribute: route prediction is performed based on the semantic feature representations to obtain the routing probabilities of each expert network in the expert network group applicable to the flight attribute. Based on the routing probabilities of each expert network and the output feature representations processed by each expert network, the target feature representation of the flight attribute is obtained. Mapping prediction is then performed based on the target feature representations to obtain the attribute values ​​of the flight attributes in the predicted flight path points. The expert network group applicable to the flight attributes includes: a shared expert network applicable to all flight attributes and a customized expert network specifically for flight attributes. Finally, based on the attribute values ​​of each flight attribute in the predicted flight path points, the attribute information of the predicted flight path points is obtained. The predicted flight path points are located after the observed flight path points. Instead of modeling it as a multi-class classification problem for each flight attribute, this approach predicts each flight attribute separately during trajectory prediction. Firstly, it performs route prediction based on semantic feature representations for each flight attribute, which helps to fully consider the unique characteristics of different flight attributes. Secondly, when predicting the attribute values ​​of each flight attribute, it relies on an expert network group containing both a shared expert network and a custom expert network. This allows the custom expert network to fully learn the unique characteristics of different flight attributes, and the shared expert network to further learn the general temporal features of the flight trajectory. Finally, using the routing probabilities of each expert network in the expert network group applicable to the flight attribute obtained during the route prediction process, along with the output feature representations processed by each expert network, the target feature representation of the flight attribute is obtained. This approach avoids the predicament of different flight attributes being forced to share the same feature representation for attribute prediction, thus improving the accuracy of mapping and predicting the attribute values ​​of the flight attributes. Therefore, it can improve the accuracy of trajectory prediction.

[0065] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0066] The description of the various embodiments above tends to emphasize the differences between the various embodiments. The similarities or similarities between them can be referred to, and for the sake of brevity, they will not be repeated here.

[0067] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus implementations described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0068] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0069] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0070] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0071] If the technical solution of this application involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

Claims

1. A trajectory prediction method, characterized in that, include: Feature embedding is performed based on the attribute information of observed flight track points to obtain embedded feature representation; wherein, the attribute information includes attribute values ​​of several flight attributes, including: longitude, latitude, altitude, heading, speed, country of origin, and aircraft type; Semantic extraction is performed based on the embedded feature representation to obtain a semantic feature representation; For each of the flight attributes: route prediction is performed based on the semantic feature representation to obtain the route probability of each expert network in the expert network group applicable to the flight attribute; and based on the route probability of each expert network and the output feature representation processed by each expert network, the target feature representation of the flight attribute is obtained; and mapping prediction is performed based on the target feature representation to obtain the attribute value of the flight attribute in the predicted waypoint; wherein, the expert network group applicable to the flight attribute includes: a shared expert network applicable to each of the flight attributes and a customized expert network specifically for the flight attribute; Based on the attribute values ​​of each flight attribute in the predicted waypoint, the attribute information of the predicted waypoint is obtained; wherein, the predicted waypoint is located after the observed waypoint, and the attribute information of the predicted waypoint is obtained by the waypoint prediction model predicting the attribute information of the observed waypoint. The waypoint prediction model is trained based on the model's main loss, which is obtained by weighting the sub-losses of each flight attribute involved in the attribute information and their loss weights. Furthermore, the loss weights of the flight attributes are negatively adjusted based on the optimization speed of the corresponding sub-losses, including: Each flight attribute is selected as the current attribute. Based on the loss weight of the flight attribute after the previous training and the sub-loss of the flight attribute in the previous training, the actual gradient norm of the flight attribute is obtained. Based on the ratio of the sub-loss of the flight attribute in this training to that in the first training, the optimization speed of the flight attribute is obtained. Based on the baseline gradient norm obtained by fusing the actual gradient norms of each flight attribute and the optimization speed of the current attribute, the expected gradient norm of the current attribute is obtained. Based on the difference between the expected gradient norm and the actual gradient norm of each flight attribute, a second regularization loss is obtained. An uncertainty loss is then obtained based on the sub-losses of each flight attribute and the noise parameters updated in the previous training iteration. Parameters are adjusted based on the second regularization loss, the uncertainty loss, and the model's main loss to obtain the network parameters updated in this training iteration of the trajectory prediction model and the noise parameters updated for each flight attribute. The rate of change of the current attribute with respect to the noise parameters is then obtained based on the noise parameters updated for the current attribute. Finally, the expected gradient norm of the current attribute is updated based on this rate of change, resulting in a new expected gradient norm for the current attribute; the new expected gradient norm is negatively correlated with the rate of change. Based on the difference between the new expected gradient norm and the actual gradient norm of the current attribute, the loss weights of the current attribute after the previous training are adjusted to obtain the loss weights of the current attribute after the current training.

2. The method according to claim 1, characterized in that, The attribute information of the predicted track points is obtained by the track prediction model predicting the attribute information of the observed track points. The training steps of the track prediction model include: In the process of predicting the attribute information of each sample observation track point in the training batch based on the trajectory prediction model, for each flight attribute: the sample routing probability and the number of times each expert network in the expert network group applicable to the flight attribute are obtained, and based on the sample routing probability and the number of times the same expert network is selected in each prediction in the training batch, the preference degree of the flight attribute for the expert network is obtained, and the load balance degree of the expert network group applicable to the flight attribute is obtained by fusing the preference degree of the flight attribute for each expert network in the expert network group applicable to the flight attribute; wherein, in the process of predicting the attribute information of any sample observation track point, the expert network with the highest sample routing probability is selected. The load balancing loss is obtained based on the load balancing degree of the expert network groups applicable to each of the aforementioned flight attributes. Based on the load balancing loss, the network parameters of the trajectory prediction model are adjusted.

3. The method according to claim 2, characterized in that, The sample observation track points are labeled with attribute information of the first sample track point, and the first sample track point is located after the sample observation track point. Before adjusting the network parameters of the track prediction model based on the load balancing loss, the method further includes: Based on the trajectory prediction model, the attribute information of the sample observation trajectory point is predicted to obtain the attribute information of the second sample trajectory point; wherein, the second sample trajectory point is located after the sample observation trajectory point; Based on the difference between the attribute information of the first sample track point and the attribute information of the second sample track point, the main loss of the model is obtained; The adjustment of network parameters of the trajectory prediction model based on the load balancing loss includes: Based on the model's main loss and the load balancing loss, the network parameters of the trajectory prediction model are adjusted.

4. The method according to claim 2, characterized in that, The method of obtaining the degree of preference of the flight attribute for the expert network based on the sample route probabilities predicted by the same expert network in each training batch and the statistical number of times it was selected includes: The first preference is obtained by averaging the sample routing probabilities predicted by the expert network in each training batch, and the second preference is obtained by the ratio of the number of times the expert network predicts that the sample route is selected in each training batch to the total number of samples in the training batch. The degree of preference of the flight attribute for the expert network is obtained by multiplying the first preference degree and the second preference degree of the same expert network.

5. The method according to claim 1, characterized in that, The process of obtaining the expected gradient norm of the current attribute based on the baseline gradient norm obtained by fusing the actual gradient norms of each of the flight attributes and the optimization speed of the current attribute includes: The baseline gradient norm is obtained by averaging the actual gradient norms of each of the aforementioned flight attributes. The baseline gradient norm is scaled based on a first scaling factor with the optimization speed of the current attribute as the base and the balancing hyperparameter as the exponent, to obtain the expected gradient norm of the current attribute.

6. The method according to claim 1, characterized in that, The step of adjusting the loss weights of the current attribute after the previous training iteration based on the difference between the expected gradient norm and the actual gradient norm of the current attribute, to obtain the adjusted loss weights of the current attribute after the current training iteration, includes: Based on the difference between the expected gradient norm and the actual gradient norm of the current attribute, the first regularization loss of the current attribute is obtained. Obtain the product of the partial derivative of the first regularization loss with respect to the loss weight and the learning rate, and use it as the adjustment range of the loss weight of the current attribute in this round of training; Based on the adjustment range of the loss weight of the current attribute in this round of training, the loss weight of the current attribute after the adjustment in the previous round of training is adjusted to obtain the loss weight of the current attribute after the adjustment in this round of training.

7. The method according to claim 1, characterized in that, The uncertainty loss is obtained by combining the sub-losses based on each flight attribute and the noise parameters updated in the previous training round, including: Based on the noise parameters obtained from the previous training update of the flight attribute, the sub-losses of the flight attribute are weighted to obtain the first sub-loss of the flight attribute, and based on the noise parameters obtained from the previous training update of the flight attribute, the second sub-loss of the flight attribute is obtained; wherein, the first sub-loss of the flight attribute is negatively correlated with the noise parameters obtained from the previous training update of the flight attribute, and the second sub-loss of the flight attribute is positively correlated with the noise parameters obtained from the previous training update of the flight attribute; The uncertainty loss is obtained by fusing the first sub-loss and the second sub-loss based on each of the aforementioned flight attributes.

8. The method according to claim 1, characterized in that, The step of updating the expected gradient norm of the current attribute based on the rate of change of the current attribute with respect to the noise parameter to obtain a new expected gradient norm of the current attribute includes: A second scaling factor for the current attribute is obtained based on the rate of change of the current attribute with respect to the noise parameter; wherein the second scaling factor is negatively correlated with the rate of change. The expected gradient norm of the current attribute is scaled based on the second scaling factor of the current attribute to obtain a new expected gradient norm of the current attribute.

9. The method according to claim 1, characterized in that, Depending on whether the different flight attributes are related, the flight attributes are divided into several attribute combinations. The attribute values ​​of each flight attribute in the same attribute combination are embedded and mapped by the same embedding matrix to obtain the embedding representation of the corresponding attribute combination. The embedding feature representation is obtained by fusing the embedding representations of the several attribute combinations.

10. The method according to claim 1, characterized in that, The feature embedding based on the attribute information of the observed track points yields the embedded feature representation, including: The first embedding matrix is ​​used to embed the attribute values ​​of longitude and latitude in the attribute information of the observed flight path point to obtain a first embedding representation. The second embedding matrix is ​​used to embed the attribute values ​​of country and aircraft type in the attribute information of the observed flight path point to obtain a second embedding representation. The third embedding matrix is ​​used to embed the attribute values ​​of heading and speed in the attribute information of the observed flight path point to obtain a third embedding representation. The fourth embedding matrix is ​​used to embed the attribute value of altitude in the attribute information of the observed flight path point to obtain a fourth embedding representation. The embedding feature representation is obtained by fusing the first embedding representation, the second embedding representation, the third embedding representation, and the fourth embedding representation; Wherein, the number of rows in the first embedding matrix is ​​the sum of the sizes of the longitude term list and the latitude term list, and the number of columns in the first embedding matrix represents the dimension of the joint embedding of longitude and latitude; the number of rows in the second embedding matrix is ​​the sum of the sizes of the country term list and the aircraft type term list, and the number of columns in the second embedding matrix represents the dimension of the joint embedding of country and aircraft type; the number of rows in the third embedding matrix is ​​the sum of the sizes of the airspeed term list and the heading term list, and the number of columns in the third embedding matrix represents the dimension of the joint embedding of airspeed and heading.

11. The method according to claim 1, characterized in that, After obtaining the attribute information of the predicted waypoint based on the attribute values ​​of each flight attribute in the predicted waypoint, the method further includes: Obtain the attribute information of the new observation track point at the current moment; The step of embedding features based on the attribute information of the observed track points to obtain the embedded feature representation is returned to obtain the attribute information of the new predicted track points.

12. The method according to any one of claims 1 to 11, characterized in that, The observed track point and the predicted track point belong to the same aircraft; And / or, where the flight attributes include longitude and latitude, the expert network group applicable to the longitude shares a custom expert network dedicated to the longitude and the latitude with the expert network group applicable to the latitude; And / or, where the flight attributes include heading and speed, the expert network group applicable to the heading and the expert network group applicable to the speed share a custom expert network dedicated to the heading and the speed. And / or, where the flight attributes include country of origin and aircraft type, the expert network group applicable to the country of origin shares a customized expert network dedicated to the country of origin and the aircraft type with the expert network group applicable to the aircraft type; And / or, where the flight attributes include altitude, the expert network group applicable to the altitude includes a custom expert network specifically for the altitude; And / or, the attribute information of the predicted track point is obtained by the track prediction model predicting the attribute information of the observed track point, and the track prediction model includes: a shared expert network applicable to each of the flight attributes and customized expert networks specifically for each of the flight attributes; And / or, the semantic extraction is performed by a feature extraction network, which includes a Transformer block; And / or, for any of the flight attributes, the route prediction is performed by a route prediction network applicable to the corresponding flight attribute; And / or, the expert network includes a fully connected layer; And / or, the target feature representation is obtained by weighted summation of the output feature representation obtained by the expert network processing the semantic feature representation using the routing probabilities of the expert network.

13. A trajectory prediction device, characterized in that, include: The feature embedding module is used to embed features based on the attribute information of observed flight track points to obtain embedded feature representations; wherein, the attribute information includes attribute values ​​of several flight attributes, including: longitude, latitude, altitude, heading, speed, country of origin, and aircraft type; The semantic extraction module is used to perform semantic extraction based on the embedded feature representation to obtain a semantic feature representation; An attribute prediction module is configured to, for each of the flight attributes: perform route prediction based on the semantic feature representation to obtain the route probability of each expert network in the expert network group applicable to the flight attribute; and obtain the target feature representation of the flight attribute based on the route probability of each expert network and the output feature representation processed by each expert network respectively on the semantic feature representation; and perform mapping prediction based on the target feature representation to obtain the attribute value of the flight attribute in the predicted waypoint; wherein, the expert network group applicable to the flight attribute includes: a shared expert network applicable to each of the flight attributes and a customized expert network specifically for the flight attribute; The trajectory acquisition module is used to obtain the attribute information of the predicted trajectory point based on the attribute values ​​of each flight attribute in the predicted trajectory point; wherein, the predicted trajectory point is located after the observed trajectory point, and the attribute information of the predicted trajectory point is obtained by the trajectory prediction model predicting the attribute information of the observed trajectory point. The trajectory prediction model is trained based on the model's main loss, which is obtained by weighting the sub-losses of each flight attribute involved in the attribute information and their loss weights. Furthermore, the loss weights of the flight attributes are negatively adjusted based on the optimization speed of the corresponding sub-losses, including: Each flight attribute is selected as the current attribute. Based on the loss weight of the flight attribute after the previous training and the sub-loss of the flight attribute in the previous training, the actual gradient norm of the flight attribute is obtained. Based on the ratio of the sub-loss of the flight attribute in this training to that in the first training, the optimization speed of the flight attribute is obtained. Based on the baseline gradient norm obtained by fusing the actual gradient norms of each flight attribute and the optimization speed of the current attribute, the expected gradient norm of the current attribute is obtained. Based on the difference between the expected gradient norm and the actual gradient norm of each flight attribute, a second regularization loss is obtained. An uncertainty loss is then obtained based on the sub-losses of each flight attribute and the noise parameters updated in the previous training iteration. Parameters are adjusted based on the second regularization loss, the uncertainty loss, and the model's main loss to obtain the network parameters updated in this training iteration of the trajectory prediction model and the noise parameters updated for each flight attribute. The rate of change of the current attribute with respect to the noise parameters is then obtained based on the noise parameters updated for the current attribute. Finally, the expected gradient norm of the current attribute is updated based on this rate of change, resulting in a new expected gradient norm for the current attribute; the new expected gradient norm is negatively correlated with the rate of change. Based on the difference between the new expected gradient norm and the actual gradient norm of the current attribute, the loss weights of the current attribute after the previous training are adjusted to obtain the loss weights of the current attribute after the current training.

14. An electronic device, characterized in that, It includes at least a memory and a processor coupled to each other, wherein the memory stores at least program instructions, and the processor is used to execute the program instructions to implement the trajectory prediction method according to any one of claims 1 to 12.

15. A computer-readable storage medium, characterized in that, The system stores program instructions that can be executed by a processor, the program instructions being used to implement the trajectory prediction method according to any one of claims 1 to 12.

Images