Track application reasoning method and system based on track pre-training large model

Abstract

The invention relates to the technical field of track analysis, in particular to a track application reasoning method and system based on a track pre-training large model, and the method comprises the steps: obtaining a high-precision open-source track data set, cleaning the data set, and constructing a self-supervised training data set; constructing a track pre-training large model for extracting high-dimensional semantic features of the track; based on the self-supervised training data set, constructing a learning loss function to pre-train the track pre-training large model; aiming at a specific track downstream application, constructing a small sample labeling training data set; and cross-domain application training and reasoning are realized based on a small sample labeling training data set. According to the method, the semantic features of the track are extracted by using the track pre-training large model, and cross-domain application training and reasoning are carried out based on cross-domain data, so that the requirement on the number of application labeling samples in special fields can be remarkably reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of track analysis, and in particular to a track application reasoning method and system based on a track pre-training large model. Background Art

[0002] Track analysis is of great significance in the fields of air traffic control, surface ship dispatching, and unmanned driving. Although in most cases, the status data, identity information, and other data of cooperative targets can be concentrated in the control center and become the basis for the control center's command decision-making, there are often phenomena such as data loss and non-cooperative target intrusion in actual applications. The status information and identity information of these targets have large noise or some attributes are missing, which brings decision-making risks to applications such as air traffic control. In order to reduce the decision-making risk, it is necessary to carry out track analysis tasks such as track smoothing, filtering, interpolation, prediction, association, classification, and identity verification on track data containing noise and missing some attributes. Existing research mainly realizes analysis tasks such as track smoothing, filtering, interpolation, and prediction based on the Kalman filtering method of a simple motion model, and realizes analysis tasks such as track classification, association, and identity verification based on data-driven end-to-end machine learning methods. In actual applications, due to the complexity of the target motion law and the number of special target labeled samples, the existing methods need to further improve the effect of track analysis. Summary of the invention

[0003] The present invention aims to solve the problem of insufficient number of annotations in special fields, and proposes a track application reasoning method and system based on a track pre-training large model, which can significantly reduce the requirement on the number of annotation samples for special field applications.

[0004] To achieve the above purpose, the technical solution adopted is:

[0005] A track application reasoning method based on a track pre-trained large model, including:

[0006] Obtain high-precision open source track data sets, clean the data sets, and construct self-supervised training data sets;

[0007] Build a large track pre-training model to extract high-dimensional semantic features of the track;

[0008] Based on the self-supervised training dataset, a learning loss function is constructed to pre-train the large track pre-training model;

[0009] Construct a small sample annotated training dataset for specific track downstream applications;

[0010] Based on small sample labeled training datasets, cross-domain application training and reasoning can be achieved.

[0011] According to the track application reasoning method based on the track pre-training large model of the present invention, further, obtaining a high-precision open source track data set refers to collecting a large amount of high-precision track data of specific types of targets through open source channels, and the track data includes high-dimensional state attributes of the track at each moment.

[0012] According to the track application reasoning method based on the track pre-training large model of the present invention, further, the data set cleaning includes: first, eliminating too short track data; eliminating track data with too many missing and too large time intervals; eliminating track data with abnormal data; and then normalizing the track attributes.

[0013] According to the track application reasoning method based on the track pre-training large model of the present invention, further, constructing a self-supervised training data set includes:

[0014] First, prepare the training data: Assume that the maximum number of tokens in the input sequence of the track pre-training large model is L, use a sliding window with a window size of L to intercept tracks from the open source track dataset, and use the [PAD] marker to fill tracks that are less than L in length;

[0015] Then, the state vectors of several tokens of the track data are randomly selected for MASK, while the timestamp parameter is not MASK. The value of [MASK] should deviate from the value interval of all self-supervised sample states; the state vector before MASK is used as the labeling information of the self-supervised sample.

[0016] According to the track application reasoning method based on the track pre-training large model of the present invention, further, the track pre-training large model includes an input layer, a multi-layer Transformer module and an output layer; the input layer uses a fully connected network to map the input track token sequence to a high-dimensional feature space; the Transformer module includes a multi-head attention module, a layer normalization layer, a feedforward neural network layer, a layer standard layer and a ResNet operation; the output layer is a fully connected network, which is used to map the track semantic feature embedding vector extracted by the multi-layer Transformer module to the parameter space of the track point attribute output.

[0017] According to the track application reasoning method based on the track pre-training large model of the present invention, further, the input of the track pre-training large model is the MASK track containing timestamp information, and the output is the probability distribution parameter of the MASK state vector; the learning loss function is constructed using the Euclidean distance as follows:

[0018]

[0019] Among them, y kj is the label of the kth token of the kth sample in the batch; is the network output vector of the kth sample; Represents the determinant of the estimated covariance matrix corresponding to the network output vector;

[0020] The last Transformer module of the multi-layer Transformer module outputs a high-dimensional semantic feature vector of the track.

[0021] According to the track application reasoning method based on the track pre-trained large model of the present invention, further, for continuous track downstream applications, the track pre-trained large model is used as a track motion model and embedded in the particle filtering algorithm to realize continuous track downstream applications; and a Euclidean distance cost function is constructed for continuous track downstream applications.

[0022] According to the track application reasoning method based on the track pre-training large model of the present invention, further, for discrete track downstream applications, a "track pre-training large model + downstream application fine-tuning network" structure is formed, and the high-dimensional semantic feature vector output by the track pre-training large model is used as the input of the downstream application fine-tuning network, wherein the downstream application fine-tuning network adopts a fully connected network, a convolutional network, an LSTM or an RNN;

[0023] Based on a small sample labeled training dataset for downstream applications of tracks, a cost function is constructed to train the track pre-trained large model and the downstream application fine-tuned network;

[0024] Use the trained trajectory pre-trained large model and downstream application fine-tuned network for downstream application reasoning.

[0025] According to the track application reasoning method based on the track pre-trained large model of the present invention, further, constructing a cost function to train the track pre-trained large model and the downstream application fine-tuning network includes: constructing a cross entropy cost function for discrete track downstream applications.

[0026] Furthermore, the present invention also provides a track application reasoning system based on a track pre-training large model, comprising a self-supervised training data set construction module, a track pre-training large model construction module, a large model pre-training module, a small sample annotated training data set construction module and a cross-domain application reasoning module, wherein:

[0027] The self-supervised training dataset construction module is used to obtain high-precision open source track datasets, clean the datasets, and construct self-supervised training datasets;

[0028] Track pre-training large model building module, used to build a track pre-training large model to extract high-dimensional semantic features of the track;

[0029] The large model pre-training module is used to construct a learning loss function to pre-train the large track pre-training model based on the self-supervised training data set;

[0030] A small sample annotated training dataset construction module is used to construct a small sample annotated training dataset for specific track downstream applications;

[0031] The cross-domain application reasoning module is used to implement cross-domain application training and reasoning based on small sample labeled training data sets.

[0032] The beneficial effects achieved by adopting the above technical solution are:

[0033] The present invention extracts high-dimensional semantic embedding vectors of tracks through self-supervised pre-training based on high-precision open source track datasets, and adapts to special field applications through cross-domain application fine-tuning. It can significantly reduce the requirements for the number of labeled samples for special field applications and solve the problem of insufficient number of annotations in special fields, thereby better realizing reasoning analysis of enriched downstream applications such as track classification, state estimation, track association, track gap filling, and track prediction in specific fields.

[0034] The present invention constructs a track analysis ecological architecture of "track pre-training large model + downstream application fine-tuning network", uses the track pre-training large model to extract the semantic features of the track, and then conducts cross-domain application reasoning based on the downstream application fine-tuning network, providing a new framework for the development of track analysis related applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0036] Figure 1 It is a flow chart of a track application reasoning method based on a track pre-training large model according to an embodiment of the present invention;

[0037] Figure 2 It is a schematic diagram of the structure of the track pre-training large model of an embodiment of the present invention;

[0038] Figure 3 is a schematic diagram of a MASK method according to an embodiment of the present invention;

[0039] Figure 4 is a schematic diagram of a downstream application fine-tuning network training according to an embodiment of the present invention;

[0040] Figure 5 It is a schematic diagram of downstream application reasoning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings of specific embodiments of the present invention to clearly and completely describe the exemplary scheme of the embodiment of the present invention. Unless otherwise defined, the technical terms or scientific terms used in the present invention should be the common meanings understood by people with ordinary skills in the field.

[0042] like Figure 1 As shown, this embodiment discloses a track application reasoning method based on a track pre-training large model, comprising:

[0043] Step S101: Obtain a high-precision open-source track dataset, clean the dataset, and construct a self-supervised training dataset.

[0044] (1) Obtaining data

[0045] A high-precision open source track dataset needs to meet three conditions: first, it must be open source, and sufficient track samples can be obtained through open source channels; second, it must be high-precision, and the observation error of the track must be much smaller than the state transfer disturbance of the track target; third, it must be adaptable, and the motion law of the target in the pre-training dataset must be as adaptable as possible to the motion law of the target in the downstream application.

[0046] Acquiring high-precision open source track data sets means collecting a large amount of high-precision track data of specific types of targets through open source channels. For example, for applications related to aircraft tracks, the ADS-B open source data set is preferred; for applications related to surface ship tracks, the AIS open source data set is preferred; for applications related to vehicle tracks, the vehicle track open source data set is preferred. Track data includes high-dimensional state attributes of the track at each moment: timestamp, longitude, latitude, altitude, speed, heading, pitch angle, aircraft model, etc.

[0047] (2) Dataset cleaning

[0048] After collecting the data, it is necessary to clean the data. Specifically, it includes: first, removing too short track data; second, removing track data with too many missing or too large time intervals; third, removing track data with abnormal data. Then normalize the track attributes. Through the above steps, we get N track data R i ,i=1,2,…,N, where t ij represents the jth timestamp of the ith track, s ij is the corresponding state vector, including position, speed, heading and other state parameters, is the length of track i.

[0049] (3) Constructing a self-supervised training dataset

[0050] First, prepare the training data: Assume that the maximum number of tokens in the input sequence of the track pre-training large model is L, use a sliding window with a window size of L to intercept tracks from the open source track dataset, and use the [PAD] marker to fill tracks with a length less than L; the formal expression is:

[0051] If the track R i length Then let the training sample B k =R i ;

[0052] If the track R i length Then the subsequent part is supplemented with [PAD] to construct training sample B k =[R i ,PAD,…,PAD], where PAD is a filling vector, usually with the same dimension as R ij Similarly, each value is a sufficiently small data.

[0053] If the track R i length Then R i Split into training samples R of length L i =[r i,v+1 ,r i,v+1 ,…,r i,v+L ],in

[0054] like Figure 3 As shown, then let the kth track B k The position, velocity and other state parameters at the jth moment Set to a fixed vector [MASK], in this embodiment, the track attribute parameters in the token with a random MASK ratio of 10%; the timestamp parameter t ij No change. The value of [MASK] should deviate from the value range of all self-supervised sample states. For example, the MASK value of position longitude and latitude and heading angle is 1000, and the value of altitude is -1000. The state vector before MASK is used as the label information y of the self-supervised sample. kj . Construct the input matrix X after MASK L .

[0055] Step S102: construct a large track pre-training model for continuous spatiotemporal sequences with unequally spaced continuous input and probability output capabilities, so as to extract high-dimensional semantic features of the track.

[0056] like Figure 2As shown in the figure, a large track pre-training model is constructed to achieve high-dimensional semantic expression of the track. The large track pre-training model includes an input layer, a multi-layer Transformer module, and an output layer; the input layer uses a fully connected network to map the input track token sequence to a high-dimensional feature space; the Transformer module includes a multi-head attention module, a layer normalization layer, a feedforward neural network layer, a layer standardization layer, and a ResNet operation; the output layer is a fully connected network, which is used to map the track semantic feature embedding vector extracted by the multi-layer Transformer module to the parameter space of the track point attribute output (probability distribution parameters of variables such as longitude, latitude, altitude, speed, heading, and pitch).

[0057] The input layer cancels the traditional onehot input structure for discrete state variables and changes it to embedding directly from the latitude and longitude continuous parameter space into the continuous parameter space of track semantics, so that the modified track pre-trained large model has the ability to process continuous spatiotemporal sequences. The input layer cancels the Position embedding link used to process equally spaced time information in the traditional Transformer architecture, and directly uses the timestamp parameter as a dimension of each input token, so that the input track vector contains unequally spaced timestamp information, and has the ability to process unequally spaced spatiotemporal sequences.

[0058] The output layer has the ability to output the probability distribution of continuous state variables (such as position variables, speed variables, etc.). The specific approach is to parameterize the probability distribution of continuous state variables and use the probability distribution parameters as regression parameters of the pre-trained large model to achieve the output of the probability distribution of the specified continuous state variables.

[0059] Step S103: adopting the time series self-supervision method, based on the self-supervision training data set, constructing a learning loss function to pre-train the track pre-training large model.

[0060] The input of the track pre-training model is the MASK track containing timestamp information. The output of the track pre-training model is the probability distribution parameters of the MASK state vector, such as the expected state and the variance of the state The cost function of the self-supervised training method is the likelihood function of the probability distribution parameters of the state variable under a given labeled state vector. The learning loss function is constructed using the Euclidean distance as follows:

[0061]

[0062] Among them, y kj is the label of the kth token of the kth sample in the batch; is the network output vector of the kth sample; Represents the determinant of the estimated covariance matrix corresponding to the network output vector.

[0063] Based on the above cost function, the large model is trained to obtain the pre-trained large model parameters, completing the pre-training of the large model.

[0064] The output of the last hidden layer of the last Transformer module of the track pre-training large model is used as the high-dimensional semantic feature vector extracted from the track to complete the embedding of the track high-dimensional semantic space.

[0065] Step S104: construct a small sample annotated training data set for specific track downstream applications.

[0066] Aiming at a variety of cross-domain track downstream applications, such as track association, track identity determination, track state estimation, multi-target tracking and other tasks, we collect track data and annotation information in specific fields, and construct small sample annotated training datasets for specific fields and specific applications.

[0067] The method for combining downstream application cross-domain application reasoning with the track pre-training large model includes: a particle filtering method based on the track pre-training large model (step S105) and a downstream application deep network method based on the track pre-training large model (steps S106-S108).

[0068] Step S105: For continuous track downstream applications such as track filtering, smoothing, interpolation, and prediction, the track pre-trained large model is used as a track motion model and embedded in the particle filter algorithm to realize continuous track downstream applications.

[0069] The particle filtering method based on the track pre-training large model refers to using the track pre-training large model to model and infer the motion law of the target, obtain the prior probability distribution information of the track, and integrate the observation information with the prior information of the motion law under the Bayesian framework. The particle filtering method is used to solve the optimal estimation of the track, and then realize the continuous track downstream applications such as track filtering, smoothing, prediction, and interpolation.

[0070] For downstream applications of continuous tracks (such as track filtering, interpolation, prediction, smoothing, etc.), a Euclidean distance cost function is constructed.

[0071] Step S106: For discrete track downstream applications such as track association, identity recognition, state estimation, intent recognition, and threat discrimination, a "track pre-training large model + downstream application fine-tuning network" structure is constructed, and the high-dimensional semantic feature vector output by the track pre-training large model is used as the input of the downstream application fine-tuning network.

[0072] like Figure 4As shown in the figure, after the high-dimensional semantic embedding output of the track pre-trained large model, the fully connected network, convolutional network, LSTM or RNN and other downstream application fine-tuning networks are connected to form the "track pre-trained large model + downstream application fine-tuning network" structure.

[0073] Step S107: Based on the small sample labeled training data set of the track downstream application, a cost function is constructed to train the track pre-trained large model and the downstream application fine-tuning network.

[0074] A cross entropy cost function is constructed for discrete track downstream applications (such as track association, identity recognition, state estimation, intent recognition, threat discrimination, etc.).

[0075] Fine-tune the large model and fine-tune the network based on a small sample annotated training dataset for downstream applications of the track. The fine-tuning training process freezes some parameters to reduce the model's degrees of freedom and avoid overfitting.

[0076] Step S108: Use the trained track pre-trained large model and the downstream application fine-tuning network to perform downstream application (such as track classification, track association, etc.) reasoning, such as Figure 5 shown.

[0077] The track pre-training large model of this embodiment uses a high-precision open source track data set, while the track downstream application can use a low-precision cross-domain small sample data set. Based on the high-precision self-supervised training data set, training can focus on training the target's motion law and reduce the impact of observation noise on training. The low-precision cross-domain downstream application can make full use of the track pre-training large model to grasp the target's motion law, focusing on solving the learning problem of small samples under the condition of observation error.

[0078] Corresponding to the above method, this embodiment proposes a track application reasoning system based on a track pre-training large model, which includes a self-supervised training data set construction module, a track pre-training large model construction module, a large model pre-training module, a small sample annotated training data set construction module and a cross-domain application reasoning module, wherein:

[0079] The self-supervised training dataset construction module is used to obtain high-precision open source track datasets, clean the datasets, and construct self-supervised training datasets;

[0080] Track pre-training large model building module, used to build a track pre-training large model to extract high-dimensional semantic features of the track;

[0081] The large model pre-training module is used to construct a learning loss function to pre-train the large track pre-training model based on the self-supervised training data set;

[0082] A small sample annotated training dataset construction module is used to construct a small sample annotated training dataset for specific track downstream applications;

[0083] The cross-domain application reasoning module is used to implement cross-domain application training and reasoning based on small sample labeled training data sets.

[0084] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0085] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0086] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.

[0087] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of combination of hardware and software.

[0088] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A track application reasoning method based on a track pre-training large model, characterized in that: Include: Obtain high-precision open source track data sets, clean the data sets, and construct self-supervised training data sets; Build a large track pre-training model to extract high-dimensional semantic features of the track; Based on the self-supervised training dataset, a learning loss function is constructed to pre-train the large track pre-training model; Construct a small sample annotated training dataset for specific track downstream applications; Based on small sample labeled training datasets, cross-domain application training and reasoning can be achieved.

2. The track application reasoning method based on the track pre-training large model according to claim 1 is characterized in that: Acquiring a high-precision open source track data set means collecting a large amount of high-precision track data of specific types of targets through open source channels. The track data includes high-dimensional state attributes of the track at each moment.

3. The track application reasoning method based on the track pre-training large model according to claim 1 is characterized in that: The cleaning of the data set includes: first, removing the track data that is too short; removing the track data with too many missing items and too large time intervals; removing the track data with abnormal data; and then normalizing the track attributes.

4. The track application reasoning method based on the track pre-training large model according to claim 1 is characterized in that: Constructing a self-supervised training dataset includes: First, prepare the training data: Assume that the maximum number of tokens in the input sequence of the track pre-training large model is L, use a sliding window with a window size of L to intercept tracks from the open source track dataset, and use the [PAD] marker to fill tracks that are less than L in length; Then, the state vectors of several tokens of the track data are randomly selected for MASK, while the timestamp parameter is not MASK. The value of [MASK] should deviate from the value interval of all self-supervised sample states; the state vector before MASK is used as the labeling information of the self-supervised sample.

5. The track application reasoning method based on the track pre-training large model according to claim 4 is characterized in that: The track pre-training model includes an input layer, a multi-layer Transformer module and an output layer; the input layer uses a fully connected network to map the input track token sequence to a high-dimensional feature space; the Transformer module includes a multi-head attention module, a layer normalization layer, a feedforward neural network layer, a layer standardization layer and a ResNet operation; the output layer is a fully connected network, which is used to map the track semantic feature embedding vector extracted by the multi-layer Transformer module to the parameter space of the track point attribute output.

6. The track application reasoning method based on the track pre-training large model according to claim 5 is characterized in that: The input of the track pre-training model is the MASK track containing timestamp information, and the output is the probability distribution parameter of the MASK state vector; the learning loss function is constructed using Euclidean distance as follows: Among them, y kj is the label of the kth token of the kth sample in the batch; is the network output vector of the kth sample; Represents the determinant of the estimated covariance matrix corresponding to the network output vector; The last Transformer module of the multi-layer Transformer module outputs a high-dimensional semantic feature vector of the track.

7. The track application reasoning method based on the track pre-training large model according to claim 1 is characterized in that: For downstream applications of continuous tracks, the large track pre-training model is used as the track motion model and embedded in the particle filter algorithm to realize downstream applications of continuous tracks. For downstream applications of continuous tracks, a Euclidean distance cost function is constructed.

8. The track application reasoning method based on the track pre-training large model according to claim 1 is characterized in that: For discrete track downstream applications, a "track pre-training large model + downstream application fine-tuning network" structure is constructed, and the high-dimensional semantic feature vector output by the track pre-training large model is used as the input of the downstream application fine-tuning network, where the downstream application fine-tuning network uses a fully connected network, convolutional network, LSTM or RNN; Based on a small sample labeled training dataset for downstream applications of tracks, a cost function is constructed to train the track pre-trained large model and the downstream application fine-tuned network; Use the trained trajectory pre-trained large model and downstream application fine-tuned network for downstream application reasoning.

9. The track application reasoning method based on the track pre-training large model according to claim 8 is characterized in that: Constructing a cost function to train the large track pre-trained model and the downstream application fine-tuning network includes: constructing a cross entropy cost function for discrete track downstream applications.

10. A track application reasoning system based on a track pre-trained large model, characterized in that: It includes a self-supervised training dataset construction module, a track pre-training large model construction module, a large model pre-training module, a small sample annotated training dataset construction module and a cross-domain application reasoning module, among which: The self-supervised training dataset construction module is used to obtain high-precision open source track datasets, clean the datasets, and construct self-supervised training datasets; Track pre-training large model building module, used to build a track pre-training large model to extract high-dimensional semantic features of the track; The large model pre-training module is used to construct a learning loss function to pre-train the large track pre-training model based on the self-supervised training data set; A small sample annotated training dataset construction module is used to construct a small sample annotated training dataset for specific track downstream applications; The cross-domain application reasoning module is used to implement cross-domain application training and reasoning based on small sample labeled training data sets.

Images