Unmanned equipment track recovery method based on generative adversarial model
Through the unmanned device trajectory recovery method based on the generative adversarial model, the BERT model is used for pre-training and design of the SeqGAN model, the uncertainty problem caused by missing points in the unmanned device trajectory is solved, the space-time dependency extraction capability and quality of the trajectory are improved, and efficient trajectory recovery is achieved.
Patent Information
- Application Number
- CN202411713507.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-27
- Publication Date
- 2025-04-15
AI Technical Summary
There are missing points in the trajectory of unmanned equipment, resulting in trajectory uncertainty. It is difficult for the prior art to effectively capture the trajectory's space-time dependence and improve the embedded representation quality of the trajectory.
The unmanned device trajectory recovery method based on the generative adversarial model is adopted, and the BERT model is used for pre-training to obtain the embedded vector representation of the road segments in the road network, and a SeqGAN model is designed to restore the missing trajectory through the adversarial training of the generator and the discriminator.
The temporal and spatial dependence extraction capability and trajectory quality of trajectory are improved, the training efficiency is enhanced, the generated trajectory has high integrity, and can meet a variety of unmanned equipment application scenarios that require high-quality trajectories.
Smart Images
Figure CN119830964A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical fields of artificial intelligence, big data, and natural language processing, and specifically relates to an unmanned equipment trajectory recovery method based on a generative adversarial model. Background Art
[0002] With the widespread use of GPS devices, a large amount of unmanned equipment trajectory data has been collected. Complete unmanned equipment trajectories can provide high-quality input for various downstream tasks. However, in actual situations, due to the limitations of equipment and signal transmission, there are often missing points in the trajectory data, which increases the uncertainty of the trajectory. Therefore, it is very important to recover the missing information in the trajectory. Summary of the invention
[0003] 1. Technical issues to be resolved
[0004] The technical problem to be solved by the present invention is: the present invention proposes an unmanned equipment trajectory recovery method, which mainly solves two problems. One is to address the problem of insufficient capture of trajectory dependencies and improve the trajectory's spatiotemporal dependency extraction capability; the other is to address the problem of insufficient embedding representation of trajectories and enhance trajectory quality and improve training efficiency.
[0005] (II) Technical solution
[0006] In order to solve the above technical problems, the present invention provides an unmanned equipment trajectory recovery method based on a generative adversarial model, comprising the following steps:
[0007] Step 1: Use the BERT model to pre-train historical trajectories to obtain the embedded vector representation of each road segment in the road network; during the trajectory recovery process, the trajectory sequence is regarded as a special text sequence, and the BERT model can learn the contextual association features between road segments;
[0008] Step 2: Design a SeqGAN model based on step 1 to restore missing trajectories
[0009] The SeqGAN model design consists of a generator G and a discriminator D. The generator G is used to restore the missing trajectory sequence to a complete trajectory. It takes the partially missing trajectory and random noise as input and outputs the complete trajectory through a recursive generation mechanism. The discriminator D is used to distinguish the generated trajectory from the real trajectory. The discriminator D adopts a convolutional neural network structure and can capture the spatiotemporal characteristics of the trajectory sequence. This step trains the generator and the discriminator. During the training process, the generator continuously generates trajectories for the discriminator to discriminate. The quality of the generator's generated trajectory and the discriminator's ability to identify trajectories are continuously improved through gradient descent.
[0010] Step 3: Based on step 2, output the trajectory recovery result that conforms to the actual road network topology.
[0011] (III) Beneficial effects
[0012] The present invention provides an unmanned equipment trajectory recovery method based on a generative adversarial model, the main advantages of which are reflected in the following aspects:
[0013] (1) A generative adversarial model-based unmanned equipment trajectory recovery method is designed to complete the trajectory recovery task. The generative adversarial model is then transferred from the image field that processes continuous data to the sequence field that processes discrete data, so that the generative adversarial model can play its powerful generative advantages in the field of unmanned equipment trajectory recovery.
[0014] (2) A partial gradient descent mechanism is designed to fix the missing positions in the generative adversarial process, so that the generative adversarial model can retain the original trajectory points and the spatiotemporal information they contain during the generation process, and generate complete trajectories of indefinite length based on the missing trajectories.
[0015] (3) The idea of pre-training is introduced to improve the embedded representation of the trajectory, and a self-supervised learning method is designed to generate low-dimensional and dense trajectory segment vectors, which improves the model training efficiency and trajectory quality. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 It is the overall framework diagram of the model of the present invention. DETAILED DESCRIPTION
[0017] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below in conjunction with the drawings and examples.
[0018] The present invention proposes a method for recovering unmanned equipment trajectories based on a generative model, which realizes the spatiotemporal orderliness of trajectories while generating unmanned equipment trajectories with high integrity, and can meet various unmanned equipment application scenarios that require high-quality trajectories, such as travel time estimation and route optimization. The problems to be solved by the present invention are: first, the problem of insufficient capture of trajectory dependencies is solved by studying the adversarial network generation process and improving the spatiotemporal dependency extraction capability of trajectories; second, the problem of insufficient embedding representation of trajectories is solved by studying pre-training methods, and enhancing trajectory quality and improving training efficiency through the embedding representation of pre-trained trajectory sections.
[0019] The unmanned equipment trajectory recovery method based on generative adversarial networks (GAN) is implemented through a pre-training module, a generative adversarial network module and an output module. The pre-training module uses the BERT model to pre-train historical trajectories and generate embedded vector representations of road sections in the road network; the generative adversarial network module uses SeqGAN to generate and restore missing trajectories and optimizes the reward function through Monte Carlo search; finally, the output module maps the generated trajectory vectors to specific road section labels to form the final road section-based trajectory sequence. Through the collaborative work of these three modules, this method can efficiently and accurately recover missing trajectory data. Each module has unique advantages and methodological characteristics, and together constitute a powerful trajectory recovery model framework. The overall framework diagram is shown below. Figure 1 shown.
[0020] Given an original trajectory x = {p1, p2, ..., p K}, where K represents the number of trajectory points. The goal of the trajectory recovery task is to generate a trajectory y = {s1, s2..., s n}, so that y contains all trajectory points in x and completes the missing trajectory points, where n (n ≥ K) represents the length of the generated trajectory.
[0021] The method comprises the following steps:
[0022] Step 1: The pre-training module uses the BERT model to pre-train historical trajectories to obtain the embedded vector representation of each road segment in the road network.
[0023] The BERT model is a deep bidirectional language model based on Transformer, which can capture long-range dependencies in text sequences through the self-attention mechanism. In the process of trajectory recovery, the present invention regards the trajectory sequence as a special text sequence, and the BERT model can effectively learn the contextual association features between road segments.
[0024] Specifically, the input of the BERT model is the historical trajectory sequence T = {t1, t2, ..., t n}, where t i represents the i-th road segment in the trajectory; the BERT model first maps the input historical trajectory sequence to a word embedding vector, and then extracts features through a multi-layer Transformer encoder; the core of the Transformer encoder is the self-attention mechanism, which can adaptively assign different weights to each road segment, thereby capturing complex dependencies between road segments; after being processed by the Transformer encoder, the context representation of each road segment can be obtained, and the specific process is shown in formula (1):
[0025] hi =BERT(t i ,T) (1)
[0026] In the formula, h i represents the context representation corresponding to the i-th road segment, BERT(t i ,T) represents the use of BERT model according to the i-th road segment t i Generate the corresponding context vector with the historical trajectory sequence T.
[0027] The context vector is the output of the middle layer of the BERT model, and the embedding vector is the final output. The embedding vector is obtained based on the context vector. The process of pre-training the historical trajectory using the BERT model, that is, the pre-training process of the trajectory embedding vector, is shown in Algorithm 1: First, initialize the vector matrix V to a zero matrix and initialize the training set T masked is an empty set; then, traverse the historical trajectory sequence T, generate missing trajectories by random masking, and add them to T masked ;Finally, using the BERT model, according to the batch size Batch size Iterative training is performed to obtain an embedding vector matrix V composed of multiple embedding vectors. These embedded vectors can effectively encode the location, topology and semantic features of the road segment, providing high-quality input representation for the subsequent generation model.
[0028] The advantage of the pre-training module is that it can use a large amount of historical trajectory data to learn the rich semantic information inside the road network in a self-supervised manner, providing high-quality initialization for the generative model. Compared with directly using one-hot or embedding representation, the embedding learned by BERT has stronger generalization ability and semantic expression ability, which helps to improve the performance of the generative model. In addition, the pre-training process of the BERT model can make full use of the advantages of GPU parallel computing, greatly improving the training efficiency.
[0029]
[0030] Step 2: The generative adversarial module uses the SeqGAN model to generate and restore missing trajectories.
[0031] The SeqGAN model is a discrete sequence generation framework based on the generative adversarial network (GAN), which can effectively overcome the gradient vanishing problem of GAN in discrete sequence generation tasks.
[0032] The SeqGAN model consists of a generator G and a discriminator D. The generator G is used to restore the missing trajectory sequence to a complete trajectory. Its core is to take the partially missing trajectory and random noise as input and output the complete trajectory through a recursive generation mechanism. The generator G can take the received missing trajectory x and noise z as input to generate a complete trajectory. The calculation process is shown in formula (2):
[0033] G T =G(x,z) (2)
[0034] Where x represents the sequence of partially missing trajectories to be restored, including known trajectory points and blank paths to be restored; z is a random vector sampled from a standard distribution (such as Gaussian distribution) to increase the diversity of generation; G T is the generated recovery trajectory.
[0035] The generator G uses a sequence generation network based on LSTM, which can recursively generate a sequence of trajectory points. At the same time, the discriminator D is used to distinguish the generated trajectory from the real trajectory. The discriminator D uses a convolutional neural network structure, which can effectively capture the spatiotemporal characteristics of the trajectory sequence. Specifically, it includes a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer extracts local features in the trajectory sequence through convolution operations and can capture the pattern in the trajectory well through different filter combinations. The pooling layer is used to reduce the dimension and concentrate features. The fully connected layer combines the extracted features and outputs a probability value representing the authenticity of the trajectory. The calculation process of the discriminator D is shown in formula (3):
[0036] D(y)=CNN(y) (3)
[0037] Where y is the input trajectory sequence, and D(y) is the probability that the sequence is identified as a true trajectory.
[0038] This step trains the generator and the discriminator. During the training process, the generator continuously generates trajectories for the discriminator to discriminate, and continuously improves the quality of the trajectory generated by the generator and the ability of the discriminator to identify the trajectory through gradient descent. In order to effectively estimate the reward value of the generated trajectory, the present invention introduces a Monte Carlo search strategy in the training process of the generator and the discriminator to solve the gradient calculation challenges in discrete sequence generation and provide a stable optimization path. Monte Carlo search samples multiple possible trajectory sequences from the generated trajectory (assuming different possibilities from time step t+1 to T) and uses the discriminator to evaluate these sampled sequences, thereby determining the reward of the generated sequence.
[0039] During the training process, Monte Carlo search is used to estimate the reward value of the generated trajectory to overcome the gradient instability problem in GAN training. Specifically, the reward function is defined by formula (4):
[0040]
[0041] In the formula, y 1:t represents the generated trajectory sequence from the first trajectory point to the tth trajectory point; MC i+1:T represents the Monte Carlo sampling sequence from time step t+1 to T; N represents the number of sampling times, which controls the number of samples extracted from the generated trajectory for estimating the reward; Concat represents the concatenation of the currently generated trajectory point with the Monte Carlo sampling sequence in the time dimension. This reward function can reflect the similarity between the generated trajectory and the true trajectory, providing an effective learning signal for the generator.
[0042] The advantage of the generative adversarial network module is that it can generate realistic trajectory recovery results through adversarial training, overcoming the problem that traditional optimization-based methods are prone to fall into local optimality. At the same time, SeqGAN's Monte Carlo search strategy also greatly improves the stability and convergence speed of training. In addition, the end-to-end training process of the generator and discriminator can automatically learn the complex mapping relationship from missing trajectories to complete trajectories without relying on manually designed features.
[0043] Step 3: The output module outputs the trajectory recovery results that conform to the actual road network topology based on steps 1 and 2.
[0044] First, the missing trajectory x and noise z are input into the trained generator to obtain the restored trajectory vector. The process is shown in formula (5):
[0045] y Generated =G Trained (x,z) (5)
[0046] In the formula, G Trained Denotes the trained generator G. In this step, the generator G uses its learned sequence generation ability to output a continuous trajectory vector, which represents the initial recovery of the missing trajectory.
[0047] Then, y is transformed into Generated Mapping to specific road segment labels, this process can be regarded as a multi-classification problem, where each road segment label represents a category. The specific process is shown in formula (6):
[0048] P = SoftMax(FC(y Generated )) (6)
[0049] Where P represents the probability distribution of each road segment being selected; FC represents the fully connected layer, which is used to compress or expand the high-dimensional trajectory vector to the dimension required by the road segment label. Through the linear transformation of the weight matrix, the trajectory vector is mapped to a new representation space; SoftMax represents the Softmax layer, which converts the output of the fully connected layer into a probability distribution. For each road segment label, the probability of its being selected is calculated, and the sum of these probabilities is 1. The calculation formula is shown in formula (7):
[0050]
[0051] where z i is the i-th element of the fully connected layer output.
[0052] By calculating the probability distribution P, we select the road segment label with the highest probability at each position to form the final road segment trajectory sequence. The calculation process is shown in formula (8):
[0053] y final =argmax(P) (8)
[0054] In the formula, argmax means finding the road segment corresponding to the position with the maximum probability in P. This operation is used to select the road segment label with the largest probability value in the probability distribution P, so that the road segment label corresponding to each trajectory point has the maximum confidence. The road segment sequence selected in this way forms a smooth and coherent recovery trajectory, which conforms to the actual road network topology.
[0055] The main function of the output module is to convert the continuous trajectory vector output by the trained generator into a discrete segment sequence. This process can be regarded as a multi-classification problem. The fully connected layer and the softmax layer are responsible for mapping the trajectory vector to the segment label space. By selecting the segment in a probability-maximizing manner, a smooth and coherent recovery trajectory can be obtained.
[0056] The advantage of this output module is that it can make full use of the road segment embedding features learned by the pre-training module (i.e., the embedding vector matrix. The embedding vector generated by the pre-training module provides high-quality input for the generator, thereby training a generator that incorporates more road network information and improving the quality of the trajectory vector generated in this step) and the high-quality trajectory vector y generated by the generation module. Generated , the two are organically combined to output the trajectory recovery result y that conforms to the actual road network topology final At the same time, the softmax output layer provides probability distribution information, which can reflect the confidence of the recovery results and provide more decision support for subsequent applications.
[0057] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A method for unmanned equipment trajectory recovery based on a generative adversarial model, characterized in that: The following steps are involved: Step 1: Use the BERT model to pre-train historical trajectories to obtain the embedded vector representation of each road segment in the road network; during the trajectory recovery process, the trajectory sequence is regarded as a special text sequence, and the BERT model can learn the contextual association features between road segments; Step 2: Design a SeqGAN model based on step 1 to restore missing trajectories The SeqGAN model design consists of a generator G and a discriminator D. The generator G is used to restore the missing trajectory sequence to a complete trajectory. It takes the partially missing trajectory and random noise as input and outputs the complete trajectory through a recursive generation mechanism. The discriminator D is used to distinguish the generated trajectory from the real trajectory. The discriminator D adopts a convolutional neural network structure and can capture the spatiotemporal characteristics of the trajectory sequence. This step trains the generator and the discriminator. During the training process, the generator continuously generates trajectories for the discriminator to discriminate. The quality of the generator's generated trajectory and the discriminator's ability to identify trajectories are continuously improved through gradient descent. Step 3: Based on step 2, output the trajectory recovery result that conforms to the actual road network topology.
2. The method according to claim 1, characterized in that In step 1, the input of the BERT model is the historical trajectory sequence T = {t1, t2, ..., t n }, where t i represents the i-th road segment in the trajectory; the BERT model first maps the input historical trajectory sequence to an embedding vector, and then extracts features through a multi-layer Transformer encoder; The Transformer encoder uses a self-attention mechanism to adaptively assign different weights to each road segment, thereby capturing the complex dependencies between road segments. After being processed by the Transformer encoder, the context representation of each road segment is obtained. The specific process is shown in formula (1): h i =BERT(t i ,T) (1) In the formula, h i represents the context representation corresponding to the i-th road segment, BERT(t i ,T) represents the use of BERT model according to the i-th road segment t i Generate the corresponding context vector with the historical trajectory sequence T; the context vector is the output of the middle layer of the BERT model, and the embedding vector is the final output, which is obtained based on the context vector.
3. The method according to claim 2, characterized in that The process of pre-training the historical trajectory using the BERT model is as follows: First, initialize the vector matrix V to a zero matrix and initialize the training set T masked is an empty set; then, traverse the historical trajectory sequence T, generate missing trajectories by random masking, and add them to T masked ;Finally, using the BERT model, according to the batch size Batch size Iterative training is performed to obtain an embedding vector matrix V composed of multiple embedding vectors. These embedded vectors can encode the location, topology and semantic features of the road section.
4. The method according to claim 1, characterized in that Step 1 uses historical trajectory data to learn the semantic information within the road network through self-supervision.
5. The method according to claim 2, characterized in that The generator G can take the missing trajectory x and noise z as input to generate a complete trajectory. The calculation process is as shown in formula (2): G T =G(x,z) (2) In the formula, x represents the sequence of partially missing trajectories to be restored, including known trajectory points and blank paths to be restored; z is a random vector sampled from a standard distribution to increase the diversity of generation; G T is the generated recovery trajectory.
6. The method according to claim 1, characterized in that The generator G uses an LSTM-based sequence generation network to recursively generate trajectory point sequences.
7. The method according to claim 2, characterized in that The discriminator D includes convolutional layers, pooling layers and fully connected layers. The convolutional layers extract local features in the trajectory sequence through convolution operations and can capture patterns in the trajectory through different filter combinations. The pooling layer is used to reduce the dimension and concentrate the features. The fully connected layer combines the extracted features and outputs a probability value indicating the authenticity of the trajectory. The calculation process of the discriminator D is shown in formula (3): D(y)=CNN(y) (3) Where y is the input trajectory sequence, and D(y) is the probability that the sequence is identified as a true trajectory.
8. The method according to claim 1, characterized in that In step 2, a Monte Carlo search strategy is introduced during the training process of the generator and the discriminator. The Monte Carlo search strategy samples multiple possible trajectory sequences from the generated trajectory and uses the discriminator to evaluate these sampled sequences, thereby determining the reward of the generated sequence.
9. The method according to claim 8, characterized in that During the training process, a Monte Carlo search strategy is used to estimate the reward value of the generated trajectory. The reward function is defined by formula (4): In the formula, y 1:t represents the generated trajectory sequence from the first trajectory point to the tth trajectory point; MC i+1:T represents the Monte Carlo sampling sequence from time step t+1 to T; N represents the number of sampling times, which controls the number of samples extracted from the generated trajectory for estimating the reward; Concat means concatenating the currently generated trajectory point with the Monte Carlo sampling sequence along the time dimension.
10. The method according to claim 2, characterized in that In step 3, the missing trajectory x and noise z are first input into the trained generator to obtain the restored trajectory vector. The process is as shown in formula (5): y Generated =G Trained (x,z) (5) In the formula, G Trained represents the trained generator G. In this step, the generator G uses its learned sequence generation ability to output a continuous trajectory vector, which represents the initial recovery of the missing trajectory; Then, y is transformed into Generated Mapping to specific road segment labels, this process is regarded as a multi-classification problem, where each road segment label represents a category. The specific process is as shown in formula (6): P=SoftMax(FC(y Generated )) (6) Where P represents the probability distribution of each road segment being selected; FC represents the fully connected layer, which is used to compress or expand the high-dimensional trajectory vector to the dimension required by the road segment label; through the linear transformation of the weight matrix, the trajectory vector is mapped to a new representation space; Sofmtax represents the Softmax layer, which converts the output of the fully connected layer into a probability distribution; for each road segment label, the probability of its being selected is calculated, and the sum of these probabilities is 1. The calculation formula is as follows: Among them, z i is the i-th element of the fully connected layer output; By calculating the probability distribution P, we select the road segment label with the highest probability at each position to form the final road segment trajectory sequence. The calculation process is as shown in formula (8): y final =argmax(P) (8) In the formula, argmax represents finding the road segment corresponding to the position with the maximum probability in P. The operation of finding is used to select the road segment label with the largest probability value in the probability distribution P, so that the road segment label corresponding to each trajectory point has the maximum confidence.