A target position estimation method based on deep learning under communication delay

Abstract

The application relates to a target position estimation method based on deep learning under communication time delay and belongs to the navigation positioning field. The application adopts a strategy of synchronization first and then positioning. In the delay distance sequence synchronization stage, a main perception node respectively processes distance information with transmission time delay from different secondary perception nodes, a Seq2Seq model is trained in a course learning and sampling mode to predict the delay distance sequence, and the predicted distance sequence is used as a distance estimation sequence in a time delay section. In the position estimation stage, the time delay distance sequence of the main perception node is combined with the distance estimation sequence of all the secondary perception nodes in the time delay section, and the combined sequence is input into an Attention-LSTM network with an attention mechanism. The mapping relationship between the distance estimation sequence and coordinates is learned by using the Attention-LSTM, so that the position of a target to be estimated at a moment is obtained. The application provides an effective basis for detecting the invasion of a target radiation source, the movement direction of the target radiation source, the location of an event and the implementation of correct decisions.

Description

Technical Field

[0001] This invention relates to a deep learning-based target position estimation method under communication latency, belonging to the field of navigation and positioning. Technical Background

[0002] Today, with increasingly complex and demanding wireless communication conditions, location has become a focal point for military and civilian applications such as air-to-ground attacks, maritime search and rescue, and target radiation source indication. Given that a single sensing node cannot accurately detect and locate targets, multi-node location methods are necessary. Existing multi-node location methods can be categorized into non-cooperative and cooperative location methods based on their collaboration mechanisms. Compared to non-cooperative location, cooperative location benefits from wireless sensor networks (WSNs) and inter-device communication technologies. Through information exchange and data fusion among multiple nodes, location accuracy can be significantly improved. Furthermore, nodes can cooperate to share communication and computational loads, distributing resources and tasks, thus reducing power consumption and costs.

[0003] Several common measurement models exist for network sensing nodes (WSNs), such as Angle of Arrival (AOA), Time of Arrival (TOA), Time Difference of Arrival (TDOA), and Received Signal Strength Indicator (RSSI). Compared to other measurement models, RSSI has lower hardware requirements and lower communication overhead, making it ideal for energy-constrained network nodes. RSSI-based cooperative localization techniques utilize multiple sensing nodes to measure the received signal strength, calculate the distance to the target at each node, and then aggregate the distance information for localization to the main sensing node to determine the target's location.

[0004] However, most traditional RSSI-based cooperative positioning techniques are only suitable for ideal zero-latency communication scenarios. In complex real-world scenarios, information exchange between different nodes will inevitably be affected by communication delays. If not handled properly, this will negatively impact the positioning results of the entire system. Therefore, designing a suitable delay compensation method to synchronize delay information and minimize the impact of communication delays on cooperative positioning is a problem that urgently needs to be solved. Summary of the Invention

[0005] To synchronize the delay information of each node and minimize the impact of communication delay on the accuracy of collaborative positioning, the main objective of this invention is to provide a deep learning-based target location estimation method under communication delay. This method employs a synchronization-then-localization strategy. In the delayed distance sequence synchronization stage, the master sensing node processes distance information with transmission delay from different secondary sensing nodes. A sequence-to-sequence (Seq2Seq) model is trained using curriculum learning and sampling to predict the delayed distance sequence, which is then used as the distance estimation sequence within the delay period. In the location estimation stage, the delay-free distance sequence of the master sensing node is combined with the distance estimation sequences of all secondary sensing nodes within the delay period and input into an LSTM (Attention-LSTM) network with an attention mechanism. The Attention-LSTM learns the mapping relationship between the distance estimation sequence and coordinates to obtain the target's position at the time to be estimated. This provides a valid basis for detecting the intrusion and movement direction of target radiation sources, identifying the location of events, and making correct decisions subsequently.

[0006] This invention enables multi-sensor node collaborative estimation of target location in scenarios considering communication delays. During the delay distance sequence synchronization stage, the Seq2Seq prediction method trained based on course learning and sampling has a smaller prediction error compared to intelligent methods such as recursive prediction and multi-model prediction, thus resulting in better positioning accuracy.

[0007] The objective of this invention is achieved through the following technical solution.

[0008] This invention discloses a deep learning-based target location estimation method under communication latency, comprising the following steps:

[0009] Step 1: Generate raw data for collaborative localization; the motion state evolution of the target radiation source in three-dimensional space follows a discrete-time Markov chain. The motion direction and velocity corresponding to each finite state in this Markov chain are called the motion state, and the motion state at the current moment is only related to the motion state at the previous moment; signal propagation follows a shadow fading model, and there are n sensing nodes collaboratively sensing the target; the first node is the master sensing node, and the rest are secondary sensing nodes; obtain the distance information between the two nodes based on their respective RSSI measurements at a certain time t.

[0010]

[0011] Where P0 is the target's average transmit power, and a is the attenuation index for a specific environment;

[0012] The primary sensing node receives and fuses distance information from each secondary sensing node, and uses RSSI-based multilateral localization to obtain the target's three-dimensional coordinates; the combination of distance information and three-dimensional coordinates is recorded as the observation data collected at that moment:

[0013]

[0014] in x represents the distance information of the nth node at time t. (t) y (t) and z (t) Let x, y, and y represent the x, y, and y coordinates of the target in its three-dimensional coordinate system, respectively. The target moves for N time moments to generate a flight trajectory. There are M different trajectories, so the original data contains observation data from (M×N) time moments.

[0015] Step 2: Considering the scenario where there is transmission delay in inter-node communication, distance information arrives in order without packet loss, if the transmission delay of a certain sensing node relative to the main sensing node is τ, and the transmitted distance data is d, then the data received by the main sensing node at time T to be located is not the real-time distance d between the sensing node and the target. (T) Instead, it's the delay distance d (T-τ) The delay distances within the delay intervals form a delay distance sequence.

[0016] To reduce the impact of asynchronous distance sequences caused by communication delay on target location estimation, a method is constructed to first synchronize the delayed distance sequence and then estimate the target location. During the delayed distance sequence synchronization stage, the main sensing node processes distance information with transmission delay from different secondary sensing nodes. The Seq2Seq model trained by course learning and sampling is used to predict the delayed distance sequence, and the predicted distance sequence is used as the distance estimation sequence within the delay segment.

[0017] Step 2.1 Input the delayed distance sequence into a Seq2Seq model with an encoder-decoder architecture. First, the encoder encodes the sequence, and after multiple recursive updates, returns the final hidden state h. t and unit state c t ;

[0018]

[0019] h t =o t ⊙tanh(c t )

[0020] Where f t i t and o t These represent the forget gate, input gate, and output gate in an LSTM unit, respectively. t-1The ⊙ symbol represents the cell state at the previous time step, and ⊙ represents matrix multiplication; then h t and c t As the initial hidden state and unit state of the decoder, during the decoding process, the input of each time step in the decoder is the predicted value or the true label of the previous time step, until the entire distance sequence is output.

[0021] In summary, the functional formalization of the Seq2Seq model is as follows:

[0022]

[0023] Where: T is the time to be located, k is the input time step, and τ is the transmission delay. The distance estimated by the primary sensing node to the secondary sensing node at time T; the distance estimates within the time delay period are arranged in chronological order to form a distance estimation sequence;

[0024] Step 2.2 The trained Seq2Seq model described in Step 2 is obtained through the following method:

[0025] First, the raw data generated in step one is split column-wise. The distance sequence measured by the i-th (i = 2, 3, ... n) secondary sensing node is divided into samples according to the time step and sliding window size. These samples are then divided into training, validation, and test sets proportionally. The samples on the training set are normalized, and after dimensionality transformation, they are input into the Seq2Seq model. The model is trained using a course learning and sampling method. In each training round, the input of each time step in the decoder selects the true label with probability p and the predicted value of the previous time step with probability 1-p, thus balancing the model convergence and error accumulation propagation issues. The probability p is changed by a planned sampling method, making p follow an exponential decay. When the number of training rounds is less than half of the total number of rounds, the p value is between 0.5 and 1, and the true label is selected as the input of the next time step. When the number of training rounds exceeds half of the total number of rounds, the p value decreases to between 0 and 0.5, and the predicted value of the previous time step is selected as the input of the next time step, thus improving the overall generalization ability of the model.

[0026] During training, mean squared error (MSE) is used as the loss function to measure the true label Y. i Compared with the predicted value Differences between them:

[0027]

[0028] Where c represents the number of samples.

[0029] After training is complete, the Seq2Seq model outputs the predicted distance sequence.

[0030] Step 3: Combine the time-delay-free distance sequence of the main sensing node with the distance estimation sequence within the time delay segment obtained in Step 2 in chronological order, and input them into the trained Attention-LSTM network. Use Attention-LSTM to learn the mapping relationship between the distance estimation sequence and the coordinates to obtain the coordinate position of the target at the time to be estimated. This provides a valid basis for detecting the intrusion and movement direction of the target radiation source, identifying the location of the event, and making correct decisions in the future.

[0031] Step 3.1 Combine the time-delay-free distance sequence of the primary sensing node with the distance estimation sequences of all secondary sensing nodes within the time delay segment in chronological order to obtain:

[0032]

[0033]

[0034] Will The input is an Attention-LSTM network; specifically, the output vector of the last hidden layer of the LSTM is first used as the input to the attention layer, passed through a fully connected layer, and then the output of the fully connected layer is normalized using a softmax function.

[0035]

[0036] The weights assigned to each hidden layer vector are obtained, and the magnitude of the weights reflects the importance of the hidden state at each time step to the output result:

[0037] S i =tanh(Wh) i +b i )

[0038] a i =softmax(S i )

[0039] Where h i For the output of the last LSTM hidden layer, W and b i S represents the weights and biases of the neurons. i The score output for each hidden layer, a i The weights are used as coefficients; then, a weighted average sum c is calculated on the hidden layer output using these weights. i :

[0040]

[0041] Finally, the output layer outputs the estimated coordinates of the target at time T; in summary, the function of the Attention-LSTM model can be formally expressed as:

[0042]

[0043] Obtaining the target's position at the time to be estimated provides a valid basis for detecting the intrusion and direction of movement of the target's radiation source, identifying the location of the event, and making correct decisions in the future.

[0044] Step 3.2 The Attention-LSTM model trained in Step 3 is obtained through the following method:

[0045] The data generated in step one is segmented by distance and coordinates, with distance having a dimension of D and coordinates having a dimension of 3. Samples are constructed using the D-dimensional distance sequence over consecutive time steps as features and the 3-dimensional coordinates of the last time step as labels. These samples are then proportionally divided into training, validation, and test sets. The samples on the training set are normalized and their dimensions are transformed before being input into the Attention-LSTM model. The Adam optimization algorithm is used to adaptively adjust the learning rate. Backpropagation is used to propagate the loss function from the output layer to the input layer using the chain rule, calculating the gradient of each parameter. In each backpropagation step, the previously calculated gradients are used to update the network parameters until the model training converges.

[0046] During training, the mean distance error (MDE) between the estimated and actual target locations is used as the loss function to measure the difference between the estimated value and the true label.

[0047]

[0048] Where m represents the number of samples. and (x) (j) ,y (j) ,z (j) ) represent the estimated coordinates and true coordinates of the target at time j, respectively.

[0049] After training is complete, the Attention-LSTM model outputs the estimated coordinates of the target at time T.

[0050] Beneficial effects:

[0051] 1. This invention discloses a deep learning-based target location estimation method under communication latency. In order to synchronize the latency information of each node and minimize the impact of communication latency on the cooperative positioning accuracy, a Seq2Seq model is used to predict the delayed distance sequence during the latency distance sequence synchronization stage, which has a smaller prediction error than intelligent methods such as recursive prediction and multi-model prediction. In the location estimation stage, an attention mechanism is used to calculate the weight of each time step, and the largest weight is added to the time step that is most likely to affect the output result, which further reduces the impact of the prediction error of the Seq2Seq model on the positioning result and improves the positioning accuracy.

[0052] 2. The present invention discloses a target position estimation method based on deep learning under communication latency. It uses course learning and sampling to train a Seq2Seq model, so that the input of each time step in the decoder selects the true label with probability p and selects the predicted value of the previous time step with probability 1-p. The probability p is changed by using planned sampling, so that when the model is sufficiently trained, the predicted value of the previous time step is used as the input of the next time step. This balances the model convergence and error accumulation and propagation problems, and also improves the generalization ability of the model. Attached Figure Description

[0053] Figure 1 A scenario diagram for multi-node collaborative positioning;

[0054] Figure 2 This is a schematic diagram of the framework of a deep learning-based target position estimation method under communication latency disclosed in this invention;

[0055] Figure 3 This is a schematic diagram illustrating the training of a Seq2Seq model using course learning and sampling methods.

[0056] Figure 4 This is a comparison chart of the positioning errors of the deep learning-based target position estimation method and other intelligent methods such as recursive distance prediction and multi-model distance prediction under different average delays, based on the communication latency disclosed in this invention. Detailed Implementation

[0057] To enable those skilled in the art to more deeply understand the implementation ideas of the present invention, the technical solutions of the present invention will be carefully and clearly described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other implementation cases obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention. The specific steps of the embodiments of the present invention will be described below in conjunction with specific scenarios.

[0058] Example:

[0059] The multi-node collaborative positioning scenario in this embodiment is as follows: Figure 1As shown, five sensing nodes (S1, S2, and S3) in a WSN (Web Sensing Network) collaborate to locate a target. Considering scenarios in complex communication environments where the location information sent by secondary sensing nodes S2 and S3 to the primary sensing node S1 has different delays, this invention provides a deep learning-based target location estimation method under communication delay to minimize the impact of communication delay caused by inter-node information interaction on collaborative positioning accuracy. This method consists of two stages. In the delay distance synchronization stage, S1 processes information from different secondary sensing nodes S1 and S2. j Distance information with transmission delay (j=2, 3, 4, 5) is used to predict the delayed distance sequence using a Seq2Seq model trained through course learning and sampling. The predicted distance sequence is then used as the distance estimation sequence within the delay segment. In the location estimation stage, the delay-free distance sequence of S1 is compared with that of S... j The distance estimation sequences within the time delay are combined and input into a trained Attention-LSTM network. The Attention-LSTM network learns the mapping relationship between the distance estimation sequences and coordinates to obtain the coordinate position of the target at the time to be estimated. This provides a valid basis for detecting the intrusion and movement direction of the target radiation source, identifying the location of the event, and making correct decisions in the future.

[0060] This embodiment discloses a deep learning-based target location estimation method under communication latency. First, it generates raw data for cooperative localization. Then, it constructs a cooperative localization framework with a latency distance sequence synchronization stage and a location estimation stage, such as... Figure 2 As shown, two deep learning models in the collaborative localization framework are trained respectively. The specific implementation steps are as follows:

[0061] Step 1: Generate raw data for collaborative localization; the 5 sensing nodes S i (i = 1, 2, 3, 4, 5) are fixed in a three-dimensional space with coordinates [250, 250, 300], [750, 250, 300], [250, 750, 300], [750, 750, 750], and [500, 500, 700], respectively. The average transmit power of target M is 50 dBm. Its motion state evolution in the three-dimensional space follows a discrete-time Markov chain. The motion direction and velocity corresponding to each finite state in this Markov chain are called the motion state, and the motion state at the current moment is only related to the motion state at the previous moment. The signal propagation in the three-dimensional space follows a shadow decay model with an attenuation index of 2.4. At time t, S i (i = 1, 2, 3, 4, 5) are all sensing M, measuring their respective received signal strength (RSSI) to obtain distance information between them:

[0062]

[0063] S1 receives and merges S j The distance information sent by (j=2, 3, 4, 5) is used to obtain the three-dimensional coordinates of M using a polygonal positioning method based on RSSI; the distance information and the three-dimensional coordinates are combined and recorded as the observation data collected at that moment:

[0064]

[0065] in x represents the distance information of the nth node at time t. (t) y (t) and z (t) These represent the x-coordinate, y-coordinate, and vertical coordinate of the target in three-dimensional coordinates, respectively; 1000 flight trajectories of the target are collected, each trajectory includes 1200 time points, and the raw data contains a total of 1.2 million observation data points.

[0066] Step 2: Considering the scenario where there is transmission delay in inter-node communication, distance information arrives in order and there is no packet loss, if the transmission delay of a certain sensing node relative to S1 is τ, and the transmitted distance data is d, then the data received by S1 at time T to be located is not the real-time distance d between the sensing node and the target in this instance. (T) Instead, it's the delay distance d (T-τ) The delay distances within the delay intervals form a delay distance sequence.

[0067] To reduce the impact of asynchronous distance sequences caused by communication delays on target location estimation, a method is constructed to first synchronize the delayed distance sequences and then estimate the target location. During the delayed distance sequence synchronization phase, S1 processes the distance sequences from S... j Distance information with transmission delay for (j=2, 3, 4, 5) is used to predict the delayed distance sequence using a Seq2Seq model trained by course learning and sampling. The predicted distance sequence is then used as the distance estimation sequence within the delay segment.

[0068] Step 2.1 After inputting the delayed distance sequence into the Seq2Seq model with encoder-decoder architecture, the encoder first encodes it, and after multiple recursive updates, returns the final hidden state h. t and unit state c t :

[0069]

[0070] h t =o t ⊙tanh(c t )

[0071] Where f t i t and o tThese represent the forget gate, input gate, and output gate in an LSTM unit, respectively. t-1 This represents the cell state at the previous time step, and ⊙ represents matrix multiplication. Then h... t and c t As the initial hidden state and unit state of the decoder, during the decoding process, the input of each time step in the decoder is the predicted value or the true label of the previous time step, until the entire distance sequence is output.

[0072] In summary, the functionality of the Seq2Seq model can be formalized as follows:

[0073]

[0074] Where: T is the time to be located, k is the input time step, and τ is the transmission delay. The distance estimates are the distance information transmitted by the S1 pair of secondary sensing nodes at time T; the distance estimates within the time delay period are arranged in chronological order to form a distance estimation sequence.

[0075] In the embodiment, S j The transmission delay τ between (j = 2, 3, 4, 5) and S1 is not necessarily the same. Their values ​​are random integers between 0 and 6, and the unit is ms.

[0076] Step 2.2 The trained Seq2Seq model described in Step 2 is obtained through the following method:

[0077] First, split the raw data generated in step one by column, S j The distance sequences measured (j = 2, 3, 4, 5) are divided into multiple samples with an input time step of 12, an output time step of 6, and a sliding window size of 1. These samples are then divided into training, validation, and test sets in a 6:2:2 ratio. In this embodiment, the batch size is set to 128, the input time step to 12, and the input dimension to 1. The samples in the training set are normalized and then converted to a dimension of (128, 12, 1) before being input into the Seq2Seq model. The learning rate is set to 0.0005, the total number of training epochs is 240, and the model is trained using a course learning and sampling method. Figure 3As shown. In each training epoch, the input to each time step in the decoder selects the true label with probability p and the predicted value from the previous time step with probability 1-p, thus balancing model convergence and error propagation issues. Planned sampling is used to change the probability p, making p follow exponential decay. When the number of training epochs is less than half of the total epochs, the p value is between 0.5 and 1, and the true label is selected as the input for the next time step. When the number of training epochs exceeds half of the total epochs, the p value decreases to between 0 and 0.5, and the predicted value from the previous time step is selected as the input for the next time step, thereby improving the overall generalization ability of the model.

[0078] During training, mean squared error (MSE) is used as the loss function to measure the true label Y. i Compared with the predicted value Differences between them:

[0079]

[0080] Where c represents the number of samples.

[0081] After training is complete, the Seq2Seq model outputs the predicted distance sequence. In this embodiment, the dimension of the predicted distance sequence output by the above model is (128, 6, 1).

[0082] Step 3: Combine the time-delay-free distance sequence of S1 with the distance estimation sequence within the time delay segment obtained in Step 2 in chronological order, and input them into the trained Attention-LSTM network. Use Attention-LSTM to learn the mapping relationship between the distance estimation sequence and the coordinates to obtain the coordinate position of the target at the time to be estimated. This provides a valid basis for detecting the intrusion and movement direction of the target radiation source, identifying the location of the event, and making correct decisions in the future.

[0083] Step 3.1 Combine the time-delay-free distance sequence of S1 with S j The distance estimation sequences (j = 2, 3, 4, 5) within the time delay interval are combined in chronological order to obtain:

[0084]

[0085]

[0086] In the embodiments, The dimensions are (128, 6, 5). The input is an Attention-LSTM network; specifically, the output vector of the last hidden layer of the LSTM is first used as the input to the attention layer, passed through a fully connected layer, and then the output of the fully connected layer is normalized using a softmax function.

[0087]

[0088] The weights assigned to each hidden layer vector are obtained, and the magnitude of the weights reflects the importance of the hidden state to the output at each time step:

[0089] S i =tanh(Wh) i +b i )

[0090] a i =softmax(S i )

[0091] Where h i For the output of the last LSTM hidden layer, W and b i S represents the weights and biases of the neurons. i The score output for each hidden layer, a i The weights are used as coefficients; then, a weighted average sum c is calculated on the hidden layer output using these weights. i :

[0092]

[0093] Finally, the output layer outputs the estimated coordinates of the target at time T; in summary, the function of the Attention-LSTM model can be formally expressed as:

[0094]

[0095] Obtaining the target's position at the time to be estimated provides a valid basis for detecting the intrusion and direction of movement of the target's radiation source, identifying the location of the event, and making correct decisions in the future.

[0096] Step 3.2 The Attention-LSTM model trained in Step 3 is obtained through the following method:

[0097] The data generated in step one is segmented by distance and coordinates, with the distance dimension being 5 and the coordinate dimension being 3. Samples are constructed using a 5-dimensional distance sequence over six consecutive time steps as features and the 3-dimensional coordinates of the last time step as labels. These samples are then divided into training, validation, and test sets in a 6:2:2 ratio. In this embodiment, the batch size is set to 128. The samples on the training set are normalized and then converted into (128, 6, 5) dimensions before being input into the Attention-LSTM model. The learning rate is set to 0.0001, and the total number of training epochs is 100. The Adam optimization algorithm is used to adaptively adjust the learning rate. Backpropagation is used, and the loss function is backpropagated from the output layer to the input layer using the chain rule, calculating the gradient of each parameter. In each backpropagation step, the previously calculated gradients are used to update the network parameters until the model converges.

[0098] During training, the mean distance error (MDE) between the estimated and actual target locations is used as the loss function to measure the difference between the estimated value and the true label.

[0099]

[0100] Where m represents the number of samples. and (x) (j) ,y (j) ,z (j) ) represent the estimated coordinates and true coordinates of the target at time j, respectively.

[0101] After training is complete, the Attention-LSTM model outputs the estimated coordinates of the target at time T.

[0102] In terms of performance comparison, such as Figure 4 As shown in the figure, the blue curve with a solid circle labeled "Seq2Seq with Sampling TrainForecast" represents the positioning error of the target position estimation method based on deep learning under different average latencies in communication latency disclosed in this embodiment. Overall, especially in high latency scenarios, the positioning accuracy of the target is better than other intelligent methods in the figure.

[0103] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A target position estimation method based on deep learning under communication latency, characterized in that: Includes the following steps, Step 1: Generate raw data for collaborative localization; the evolution of the target radiation source's motion state in three-dimensional space follows a discrete-time Markov chain. The direction and velocity of motion corresponding to each finite state in this Markov chain are called the motion state, and the motion state at the current moment depends only on the motion state at the previous moment; signal propagation follows a shadowing decay model, and existing... Several sensing nodes collaborate to sense the target; the first node is the primary sensing node, and the rest are secondary sensing nodes; based on a certain... The distance information between the two parties is obtained by constantly measuring their respective received signal strength (RSSI): in It is the target's average transmission power. It is a decay index for a specific environment; The primary sensing node receives and fuses distance information from each secondary sensing node, and uses RSSI-based multilateral localization to obtain the target's three-dimensional coordinates; the combination of distance information and three-dimensional coordinates is recorded as the observation data collected at that moment: in Indicates the first Each node Distance information at any given time , and These represent the x-coordinate, y-coordinate, and vertical coordinate in the target's three-dimensional coordinate system, respectively. The target's motion generates a flight trajectory over N time moments, and there are M different trajectories. Therefore, the original data contains observation data from M×N time moments. Step 2: Considering the scenario where there is transmission delay in inter-node communication, distance information arrives in order, and there is no packet loss, the transmission delay of a certain sensing node relative to the master sensing node is... The transmitted distance data is Therefore, the data received by the primary sensing node at time T is not the real-time distance between the sensing node and the target at that time. Instead, it's the delay distance. The delay distances within the delay intervals form a delay distance sequence. To reduce the impact of asynchronous distance sequences caused by communication delay on target location estimation, a method is constructed to first synchronize the delayed distance sequence and then estimate the target location. During the delayed distance sequence synchronization stage, the main sensing node processes distance information with transmission delay from different secondary sensing nodes. The Seq2Seq model trained by course learning and sampling is used to predict the delayed distance sequence, and the predicted distance sequence is used as the distance estimation sequence within the delay segment. The specific implementation method of step two is as follows: Step 2.1 Input the delayed distance sequence into a Seq2Seq model with an encoder-decoder architecture. First, the encoder encodes the sequence, and after multiple recursive updates, returns the final hidden state. and unit state ; in , and These represent the forget gate, input gate, and output gate in an LSTM unit, respectively. This indicates the cell state at the previous time step. Represents matrix multiplication; then and As the initial hidden state and unit state of the decoder, during the decoding process, the input of each time step in the decoder is the predicted value or the true label of the previous time step, until the entire distance sequence is output. In summary, the functional formalization of the Seq2Seq model is as follows: = Where: T is the time to be located, k is the input time step, and τ is the transmission delay. This is an estimate of the distance transmitted by the secondary sensing node at time T; the distance estimates within the time delay period are arranged in chronological order to form a distance estimation sequence; Step 2.2 The trained Seq2Seq model described in Step 2 is obtained through the following method: First, split the raw data generated in step one by column. The distance sequence measured by each sensing node is divided into samples according to the time step and the sliding window size. These samples are then divided proportionally into training, validation, and test sets. The samples in the training set are normalized, their dimensions transformed, and then input into the Seq2Seq model. The model is trained using a course learning and sampling method. In each training epoch, the input to each time step in the decoder selects the true label with probability p and the predicted value from the previous time step with probability 1-p, thus balancing model convergence and error propagation. Planned sampling is used to change the probability p, making p follow exponential decay. When the number of training epochs is less than half of the total epochs, the p value is between 0.5 and 1, and the true label is selected as the input for the next time step. When the number of training epochs exceeds half of the total epochs, the p value decreases to between 0 and 0.5, and the predicted value from the previous time step is selected as the input for the next time step. During training, mean squared error (MSE) is used as the loss function to measure the true label. Compared with the predicted value Differences between them: Where c represents the number of samples; After training is complete, the Seq2Seq model outputs the predicted distance sequence; Step 3: Combine the time-delay-free distance sequence of the master sensing node with the distance estimation sequence within the time delay segment obtained in Step 2 in chronological order, and input it into the trained LSTM (Attention-LSTM) network with attention mechanism. Use Attention-LSTM to learn the mapping relationship between the distance estimation sequence and the coordinates to obtain the coordinate position of the target at the time to be estimated.

2. The target position estimation method based on deep learning under communication latency as described in claim 1, characterized in that: The specific implementation method of step three is as follows: Step 3.1 Combine the time-delay-free distance sequence of the primary sensing node with the distance estimation sequences of all secondary sensing nodes within the time delay segment in chronological order to obtain: , Will The input is an Attention-LSTM network; specifically, the output vector of the last hidden layer of the LSTM is first used as the input to the attention layer, passed through a fully connected layer, and then the output of the fully connected layer is normalized using a softmax function. The weights assigned to each hidden layer vector are obtained, and the magnitude of the weights reflects the importance of the hidden state at each time step to the output result: in This is the output of the last LSTM hidden layer. and For the weights and biases of neurons, The score output for each hidden layer. These are the weighting coefficients; then, a weighted average sum is calculated on the hidden layer output using these weighting coefficients. : Finally, the output layer outputs the estimated coordinates of the target at time T; in summary, the function of the Attention-LSTM model can be formally expressed as: Obtaining the target's position at the time to be estimated provides a valid basis for detecting the intrusion and direction of movement of the target's radiation source, identifying the location of the event, and making correct decisions in the future. Step 3.2 The Attention-LSTM model trained in Step 3 is obtained through the following method: The data generated in step one is segmented by distance and coordinates, with distance having a dimension of D and coordinates having a dimension of 3. Samples are constructed using the D-dimensional distance sequence over consecutive time steps as features and the 3-dimensional coordinates of the last time step as labels. These samples are then proportionally divided into training, validation, and test sets. The samples on the training set are normalized and their dimensions are transformed before being input into the Attention-LSTM model. The Adam optimization algorithm is used to adaptively adjust the learning rate. Backpropagation is used to propagate the loss function from the output layer to the input layer using the chain rule, calculating the gradient of each parameter. In each backpropagation step, the previously calculated gradients are used to update the network parameters until the model training converges. During training, the mean distance error (MDE) between the estimated and actual target locations is used as the loss function to measure the difference between the estimated value and the true label. Where m represents the number of samples. and Let represent the estimated coordinates and true coordinates of the target at time j, respectively; After training is complete, the Attention-LSTM model outputs the estimated coordinates of the target at time T.

Images