Target system and method for locating point of impact based on deep learning time difference estimation

Abstract

This invention relates to a target system and a projectile impact point positioning method based on deep learning time difference estimation. The target system based on deep learning time difference estimation includes a target, piezoelectric sensors, signal processing circuits, an analog-to-digital converter, a deep learning time difference estimation module, and a coordinate positioning module. The projectile impact point positioning method includes a signal acquisition step, a signal preprocessing step, a deep learning time difference estimation step, and a coordinate calculation step. It directly outputs the wave arrival time difference estimate from the end-to-end of the three-channel shock wave waveform, avoiding the accumulated errors caused by traditional multi-step processing. The entire positioning method only requires three piezoelectric sensors to achieve high-precision positioning of the projectile impact point, significantly reducing hardware costs and system complexity, and also improving the accuracy of wave arrival time difference estimation.

Description

Technical Field

[0001] This invention relates to the technical field of target systems and bullet impact point positioning methods based on deep learning time difference estimation, and in particular to a target system and bullet impact point positioning method based on deep learning time difference estimation. Background Technology

[0002] Traditional shooting training and competitions often use paper targets as the interface for presenting bullet impact points due to cost considerations. However, paper targets have many limitations in actual use.

[0003] First, paper targets rely entirely on manual interpretation of bullet hole locations, which is not only time-consuming but also susceptible to factors such as observation angle, paper damage, or overlapping holes caused by multiple shots, resulting in inaccurate bullet impact point determination. Second, paper targets are prone to cracking or deformation due to continuous bullet penetration, requiring frequent replacements, which not only increases material costs but also interrupts the shooting training process, reducing efficiency.

[0004] Later, electronic targets with automatic target reporting appeared, but they generally use piezoelectric sensors, which have the following shortcomings: 1. When using traditional signal processing methods such as cross-correlation or STA / LTA to extract time of arrival, the accuracy of time of arrival estimation decreases significantly under conditions of low signal-to-noise ratio, strong environmental noise, and complex interference, leading to increased positioning error.

[0005] 2. Poor adaptability: Traditional methods rely on manually set thresholds and characteristic parameters, which are not adaptable to different target materials, different types of ammunition and environmental changes, making it difficult to achieve generalized application.

[0006] 3. Numerous sensors and complex systems: Existing solutions often use 4, 6 or even more sensors, which increases hardware costs and system complexity.

[0007] Therefore, in this patent application, the applicant has carefully studied a target system and a bullet impact point localization method based on deep learning time difference estimation to solve the above problems. Summary of the Invention

[0008] The present invention addresses the shortcomings of the existing technology and aims to provide a target system and projectile impact point positioning method based on deep learning time difference estimation. It directly outputs the time difference of arrival estimation from the end-to-end of the three-channel shock wave waveform, avoiding the cumulative error caused by traditional multi-step processing. The entire positioning method only requires three piezoelectric sensors to achieve high-precision positioning of the projectile impact point, significantly reducing hardware costs and system complexity, and improving the accuracy of time difference of arrival estimation.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A target system based on deep learning time difference estimation includes: The target has a target face and a back face that are positioned opposite to each other. Three piezoelectric sensors are arranged in an equilateral triangle on the back side. The signal processing circuit consists of three circuits, each connected to a piezoelectric sensor to receive and process signals. There are three analog-to-digital converters. Each analog-to-digital converter is connected to the output of a signal processing circuit to acquire the voltage value output by the connected signal processing circuit through a threshold method and realize analog-to-digital conversion. The deep learning time difference estimation module has a pre-trained deep learning model, which is connected to the output of all analog-to-digital converters to preprocess the collected three-channel data and then input it into the pre-trained deep learning model to directly output the time difference of arrival. The coordinate positioning module is connected to the deep learning time difference estimation module to calculate the impact point coordinates based on the time difference output by the deep learning model and the known coordinates of the sensor.

[0010] As a preferred embodiment, the signal processing circuit includes a charge amplifier, a voltage amplifier, and a level conversion circuit connected in sequence. The charge amplifier is connected to a corresponding piezoelectric sensor, and the output terminal of the level conversion circuit is the output terminal of the signal processing circuit.

[0011] As a preferred embodiment, based on the target system based on deep learning time difference estimation, three piezoelectric sensors are defined in a counterclockwise direction: piezoelectric sensor S0, piezoelectric sensor S1, and piezoelectric sensor S2, wherein piezoelectric sensor S0 is located at the lower left corner of an equilateral triangle. The method includes the following steps: (1) Signal acquisition steps: After the piezoelectric sensor S0, piezoelectric sensor S1 and piezoelectric sensor S2 sense the shock wave signal, they output to the corresponding signal processing circuit for processing. After processing, the signal processing circuit sends the signal to the corresponding analog-to-digital converter. Each analog-to-digital converter acquires the voltage value output by the corresponding signal processing circuit through the threshold method and realizes analog-to-digital conversion, that is, obtains the waveform data of the corresponding channel. The threshold method includes acquiring waveform data of a preset time length when the signal amplitude exceeds the preset threshold. (2) Signal preprocessing steps: preprocess the waveform data of the three corresponding channels to obtain the preprocessed three-channel waveform data; (3) Deep learning time difference estimation steps: Input the preprocessed three-channel waveform data into the pre-trained deep learning model. The pre-trained deep learning model directly outputs the time difference of arrival of the two waves: Time difference of arrival dt1 = Time difference between the arrival of the shock wave at piezoelectric sensor S1 and the arrival at piezoelectric sensor S0; Time difference of arrival dt2 = Time difference between the arrival of the shock wave at piezoelectric sensor S2 and the arrival at piezoelectric sensor S0; (4) Coordinate calculation steps: Given the spatial coordinates of piezoelectric sensor S0, piezoelectric sensor S2, and the propagation velocity v of the shock wave in the target material, establish a set of equations based on the time difference positioning principle: Let a Cartesian coordinate system be established with the lower left corner of the back as the origin, and the coordinates of the impact point P be (x, y). The spatial coordinates of piezoelectric sensor S0, piezoelectric sensor S1, and piezoelectric sensor S2 are (x0, y0), (x1, y1), and (x2, y2), respectively. Then we have: =v·dt1; =v·dt2; By solving the two equations above, the precise coordinates of the impact point P are obtained and transmitted to the target reporting terminal for display.

[0012] As a preferred embodiment, the method for preprocessing the waveform data of the three corresponding channels includes performing the following sequentially: Clipping: Obtain the index number when the signal amplitude of the three channels exceeds the threshold, take the maximum value as the data endpoint, and retain the first half of the data for a preset time length; DC removal: Eliminates DC bias components in the data; Filtering and noise reduction: Use bandpass filters or wavelet filters to remove noise interference; Normalization: Normalizes the signal amplitude to the input range of a pre-trained deep learning model.

[0013] As a preferred embodiment, the method for processing the piezoelectric sensors S0, S1, and S2 after they respectively sense the shock wave signal to the corresponding signal processing circuit includes: Piezoelectric sensors S0, S1, and S2 all convert the shock wave signal into a charge signal, which is then converted into a voltage signal by the charge amplifier of the signal processing circuit, further amplified by the voltage amplifier of the signal processing circuit, and adjusted to a suitable level range by the level conversion circuit of the signal processing circuit.

[0014] As a preferred approach, training methods for deep learning models include: Multi-scenario live-fire data acquisition steps: Conduct a large number of live-fire experiments under different target materials, different ammunition types and different environmental noise conditions, and acquire waveform data from the three-channel piezoelectric sensor; Steps for obtaining the real time difference label: Use a high-precision assisted positioning system to record the real bullet impact point coordinates (x, y) for each shot, and combine the known coordinates of the three sensors and the propagation speed v of the shock wave in the target to back-calculate the corresponding real time difference of arrival (dt1_label, dt2_label), and use it as the supervised learning label; Data preprocessing steps: The collected three-channel waveform data are subjected to DC removal, bandpass filtering, and amplitude normalization. Waveform data are extracted according to a fixed time window, paired with supervised learning labels to form a training sample set, and divided into training set and validation set according to the ratio. Steps to build a deep learning model: Build a deep learning model using convolutional neural networks, long short-term memory networks, Transformer networks, or hybrid architectures. Take three-channel waveform data as input and output the predicted time difference of arrival (dt1) and time difference of arrival (dt2). Model training and optimization steps: Using mean squared error or mean absolute error as the loss function, measure the error between the model's predicted (dt1_pred, dt2_pred) and the true label (dt1_label, dt2_label). Use the backpropagation algorithm and Adam optimizer to iteratively update the network weights until the model converges on the validation set. Save the trained model weight file.

[0015] Compared with existing technologies, this invention has significant advantages and beneficial effects. Specifically, it mainly requires only three piezoelectric sensors arranged in an equilateral triangle, combined with a deep learning model, to directly output the time difference of arrival (TDOA) estimate from the end-to-end of the three-channel shock wave waveform. This avoids the cumulative error caused by traditional multi-step processing. The entire positioning method only requires three piezoelectric sensors to achieve high-precision positioning of the impact point, significantly reducing hardware costs and system complexity. Moreover, by directly outputting the TDOA end-to-end with the deep learning model, the accuracy of the TDOA estimation is improved. Secondly, by learning a large number of shock wave waveform features under different signal-to-noise ratios, target materials, and ammunition types, deep learning models can achieve robust time difference estimation in noisy and complex environments, thereby enhancing noise resistance and generalization ability. Furthermore, a threshold method is used to control the analog-to-digital converter to acquire waveform data of 400μs in length, minimizing the amount of data acquisition while ensuring complete capture of the shock wave signal, thus adapting to the time window requirements of the deep learning model input.

[0016] To more clearly illustrate the structural features and effects of the present invention, a detailed description is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the input layer structure of a preferred embodiment of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, a target system based on deep learning time difference estimation includes: The target has a front and a back side that are positioned opposite each other. In this embodiment, the target is a 10mm steel plate shooting target.

[0020] Three piezoelectric sensors are arranged in an equilateral triangle on the back side.

[0021] The signal processing circuit comprises three circuits, each connected to a piezoelectric sensor to receive and process signals. In this embodiment, the signal processing circuit includes a charge amplifier, a voltage amplifier, and a level conversion circuit connected in sequence, with the charge amplifier connected to a corresponding piezoelectric sensor.

[0022] There are three analog-to-digital converters (ADCs). Each ADC is connected to the output of a signal processing circuit to acquire the voltage value output by the connected signal processing circuit using a threshold method and to achieve analog-to-digital conversion. In this embodiment, the output of the level conversion circuit is the output of the signal processing circuit.

[0023] The deep learning time difference estimation module has a pre-trained deep learning model. It is connected to the output of all analog-to-digital converters to preprocess the collected three-channel data and then input it into the pre-trained deep learning model to directly output the time difference of arrival.

[0024] The coordinate positioning module is connected to the deep learning time difference estimation module to calculate the impact point coordinates based on the time difference output by the deep learning model and the known coordinates of the sensor.

[0025] A bullet impact point positioning method, characterized in that: it is based on the target system based on deep learning time difference estimation as described in claim 1 or 2, and defines three piezoelectric sensors, namely piezoelectric sensor S0, piezoelectric sensor S1 and piezoelectric sensor S2, in a counterclockwise direction, wherein piezoelectric sensor S0 is located at the lower left corner of an equilateral triangle; The method includes the following steps: (1) Signal acquisition steps: After piezoelectric sensors S0, S1, and S2 respectively sense the shock wave signal, they output the signal to the corresponding signal processing circuit for processing. In this embodiment, the method for piezoelectric sensors S0, S1, and S2 to output the signal to the corresponding signal processing circuit for processing after sensing the shock wave signal includes: Piezoelectric sensors S0, S1, and S2 all convert the shock wave signal into a charge signal, which is then converted into a voltage signal by the charge amplifier of the signal processing circuit, further amplified by the voltage amplifier of the signal processing circuit, and adjusted to a suitable level range by the level conversion circuit of the signal processing circuit.

[0026] After processing by the signal processing circuit, the signal is sent to the corresponding analog-to-digital converter (ADC). Each ADC acquires the voltage value output by the corresponding signal processing circuit using a threshold method and performs analog-to-digital conversion, thus obtaining the waveform data of the corresponding channel. The threshold method includes acquiring waveform data for a preset time length when the signal amplitude exceeds a preset threshold. Preferably, the preset time is 400 microseconds.

[0027] (2) Signal preprocessing step: Preprocess the waveform data of the three corresponding channels to obtain preprocessed three-channel waveform data. In this embodiment, the method for preprocessing the waveform data of the three corresponding channels includes performing the following steps sequentially: Clipping: Obtain the index number when the signal amplitude of the three channels exceeds the threshold, take the maximum value as the data endpoint, and retain the first half of the data for a preset time length; DC removal: Eliminates DC bias components in the data; Filtering and noise reduction: Use bandpass filters or wavelet filters to remove noise interference; Normalization: Normalizes the signal amplitude to the input range of a pre-trained deep learning model.

[0028] Therefore, to meet the input requirements of deep learning models, a specific preprocessing procedure is needed to remove DC, filter and denoise the three-channel shock wave signal, and normalize it.

[0029] In this embodiment, the training method for the deep learning model includes: Multi-scenario live-fire data acquisition steps: Conduct a large number of live-fire experiments under different target materials, different ammunition types and different environmental noise conditions, and acquire waveform data from the three-channel piezoelectric sensor; Steps for obtaining the real time difference label: Use a high-precision assisted positioning system to record the real bullet impact point coordinates (x, y) for each shot, and combine the known coordinates of the three sensors and the propagation speed v of the shock wave in the target to back-calculate the corresponding real time difference of arrival (dt1_label, dt2_label), and use it as a supervised learning label; preferably, the high-precision assisted positioning system is a high-speed camera system or an infrared target system.

[0030] Data preprocessing steps: The collected three-channel waveform data are subjected to DC removal, bandpass filtering, and amplitude normalization. The waveform data is extracted according to a fixed time window, paired with supervised learning labels to form a training sample set, and divided into training set and validation set according to the ratio; preferably, the fixed time window is 400μs.

[0031] Steps to build a deep learning model: Build a deep learning model using convolutional neural networks, long short-term memory networks, Transformer networks, or hybrid architectures. Take three-channel waveform data as input and output the predicted time difference of arrival (dt1) and time difference of arrival (dt2). Model training and optimization steps: Using mean squared error or mean absolute error as the loss function, measure the error between the model's predicted (dt1_pred, dt2_pred) and the true label (dt1_label, dt2_label). Use the backpropagation algorithm and Adam optimizer to iteratively update the network weights until the model converges on the validation set. Save the trained model weight file.

[0032] (3) Deep learning time difference estimation steps: Input the preprocessed three-channel waveform data into the pre-trained deep learning model. The pre-trained deep learning model directly outputs the time difference of arrival of the two waves: Time difference of arrival dt1 = Time difference between the arrival of the shock wave at piezoelectric sensor S1 and the arrival at piezoelectric sensor S0; The time difference of arrival dt2 is the time difference between the arrival of the shock wave at the piezoelectric sensor S2 and the arrival at the piezoelectric sensor S0.

[0033] In this embodiment, as shown in the figure, the principle by which the pre-trained deep learning model directly outputs the time difference of arrival (TDOA) of two waves is as follows: (I) Input Layer (Waveform Input): The model input is a preprocessed three-channel digital waveform matrix with a shape of 3×N. The first dimension corresponds to the number of channels of piezoelectric sensors S0, S1, and S2, respectively, and the second dimension corresponds to the number of discrete sampling points within a 200μs time window (e.g., N=2000). The waveforms of the three channels are synchronously aligned on the same time axis. As can be clearly seen in the figure, due to the different positions of the impact points, there is a slight but identifiable phase difference (i.e., the time difference to be estimated) in the time when the shock wave arrives at the three piezoelectric sensors.

[0034] (II) Feature Extraction Layer (Internal Processing of Deep Learning Model): The model input undergoes feature transformation sequentially through the following neural network layers: One-dimensional convolutional layer (1D-CNN): The convolutional kernel slides in the time dimension to automatically extract local abrupt change features in each waveform (such as the shock wave start point, peak slope, and steepness of the rising edge) and generate multi-channel feature maps; Pooling layer: Downsamples multi-channel feature maps, compresses data dimensionality, and enhances the translation invariance of features; Temporal Dependency Extraction Layer: The local mutation feature sequence output by the one-dimensional convolutional layer is input into the recurrent neural network (such as the Long Short-Term Memory Network LSTM or Gated Recurrent Unit GRU) or the Transformer coding layer to capture the long-range temporal dependencies between and within the three channels. The core function of this layer is to automatically compare the small phase shifts between the three waveforms. Flattening layer and fully connected layer: The high-dimensional feature vector output by the temporal dependency extraction layer is flattened into a one-dimensional vector, and then nonlinearly mapped through several layers of fully connected network to deeply fuse the temporal features into regression features that are strongly correlated with time difference.

[0035] (III) Output Layer (Time Difference Output): This is a fully connected regression layer containing two linear neuron nodes. It outputs two floating-point predicted values ​​directly without an activation function (or with a linear activation function). dt1 (Time difference of arrival of piezoelectric sensor S1 in channel S1 relative to piezoelectric sensor S0 in channel S0, in microseconds). dt2 (Time difference of arrival of channel S2, i.e., piezoelectric sensor S2, relative to channel S0, i.e., piezoelectric sensor S0, in microseconds).

[0036] The model achieves a one-step mapping from "original multi-channel waveforms" to "key time difference parameters" through end-to-end supervised learning, avoiding the multi-step cumulative error caused by "picking absolute times for each channel and then subtracting them" in traditional methods.

[0037] (4) Coordinate calculation steps: Given the spatial coordinates of piezoelectric sensor S0, piezoelectric sensor S2, and the propagation velocity v of the shock wave in the target material, establish a set of equations based on the time difference positioning principle: Let a Cartesian coordinate system be established with the lower left corner of the back as the origin, and the coordinates of the impact point P be (x, y). The spatial coordinates of piezoelectric sensor S0, piezoelectric sensor S1, and piezoelectric sensor S2 are (x0, y0), (x1, y1), and (x2, y2), respectively. Then we have: =v·dt1; =v·dt2; By solving the two equations above, the precise coordinates of the impact point P are obtained and transmitted to the target reporting terminal for display.

[0038] The key design feature of this invention is that it requires only three piezoelectric sensors arranged in an equilateral triangle, along with a deep learning model, to directly output the time difference of arrival (TDOA) estimate from the end-to-end of the three-channel shock wave waveform. This avoids the cumulative errors caused by traditional multi-step processing. The entire positioning method only requires three piezoelectric sensors to achieve high-precision positioning of the impact point, significantly reducing hardware costs and system complexity. Moreover, by directly outputting the TDOA end-to-end with the deep learning model, the accuracy of the TDOA estimation is improved. Secondly, by learning a large number of shock wave waveform features under different signal-to-noise ratios, target materials, and ammunition types, deep learning models can achieve robust time difference estimation in noisy and complex environments, thereby enhancing noise resistance and generalization ability. Furthermore, a threshold method is used to control the analog-to-digital converter to acquire waveform data of 400μs in length, minimizing the amount of data acquisition while ensuring complete capture of the shock wave signal, thus adapting to the time window requirements of the deep learning model input.

[0039] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the technical scope of the present invention. Therefore, any minor modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A target system based on deep learning time difference estimation, characterized in that: Including The target has a target face and a back face that are positioned opposite to each other. Three piezoelectric sensors are arranged in an equilateral triangle on the back side. The signal processing circuit consists of three circuits, each connected to a piezoelectric sensor to receive and process signals. There are three analog-to-digital converters. Each analog-to-digital converter is connected to the output of a signal processing circuit to acquire the voltage value output by the connected signal processing circuit through a threshold method and realize analog-to-digital conversion. The deep learning time difference estimation module has a pre-trained deep learning model, which is connected to the output of all analog-to-digital converters to preprocess the collected three-channel data and then input it into the pre-trained deep learning model to directly output the time difference of arrival. The coordinate positioning module is connected to the deep learning time difference estimation module to calculate the impact point coordinates based on the time difference output by the deep learning model and the known coordinates of the sensor.

2. The target system based on deep learning time difference estimation according to claim 1, characterized in that: The signal processing circuit includes a charge amplifier, a voltage amplifier, and a level conversion circuit connected in sequence. The charge amplifier is connected to a corresponding piezoelectric sensor, and the output terminal of the level conversion circuit is the output terminal of the signal processing circuit.

3. A method for locating the point of impact, characterized in that: Based on the target system based on deep learning time difference estimation as described in claim 1 or 2, it defines three piezoelectric sensors, namely piezoelectric sensor S0, piezoelectric sensor S1 and piezoelectric sensor S2, in a counterclockwise direction, wherein piezoelectric sensor S0 is located at the lower left corner of an equilateral triangle; The method includes the following steps: (1) Signal acquisition steps: After the piezoelectric sensor S0, piezoelectric sensor S1 and piezoelectric sensor S2 sense the shock wave signal, they output to the corresponding signal processing circuit for processing. After processing, the signal processing circuit sends the signal to the corresponding analog-to-digital converter. Each analog-to-digital converter acquires the voltage value output by the corresponding signal processing circuit through the threshold method and realizes analog-to-digital conversion, that is, obtains the waveform data of the corresponding channel. The threshold method includes acquiring waveform data of a preset time length when the signal amplitude exceeds the preset threshold. (2) Signal preprocessing steps: preprocess the waveform data of the three corresponding channels to obtain the preprocessed three-channel waveform data; (3) Deep learning time difference estimation steps: Input the preprocessed three-channel waveform data into the pre-trained deep learning model. The pre-trained deep learning model directly outputs the time difference of arrival of the two waves: Time difference of arrival dt1 = Time difference between the arrival of the shock wave at piezoelectric sensor S1 and the arrival at piezoelectric sensor S0; Time difference of arrival dt2 = Time difference between the arrival of the shock wave at piezoelectric sensor S2 and the arrival at piezoelectric sensor S0; (4) Coordinate calculation steps: Given the spatial coordinates of piezoelectric sensor S0, piezoelectric sensor S2, and the propagation velocity v of the shock wave in the target material, establish a set of equations based on the time difference positioning principle: Let a Cartesian coordinate system be established with the lower left corner of the back as the origin, and the coordinates of the impact point P be (x, y). The spatial coordinates of piezoelectric sensor S0, piezoelectric sensor S1, and piezoelectric sensor S2 are (x0, y0), (x1, y1), and (x2, y2), respectively. Then we have: =v·dt1; =v·dt2; By solving the two equations above, the precise coordinates of the impact point P are obtained and transmitted to the target reporting terminal for display.

4. The bullet impact point positioning method according to claim 3, characterized in that: The method for preprocessing the waveform data of the three corresponding channels includes performing the following steps sequentially: Crop: Obtain the index number when the signal amplitude of the three channels exceeds the threshold, take the maximum value as the data endpoint, and retain the first half of the data for a preset time length; DC removal: Eliminates DC bias components in the data; Filtering and noise reduction: Use bandpass filters or wavelet filters to remove noise interference; Normalization: Normalizes the signal amplitude to the input range of a pre-trained deep learning model.

5. The bullet impact point positioning method according to claim 3, characterized in that: The method for processing the shock wave signal after the piezoelectric sensors S0, S1, and S2 respectively sense it includes the following: Piezoelectric sensors S0, S1, and S2 all convert the shock wave signal into a charge signal, which is then converted into a voltage signal by the charge amplifier of the signal processing circuit, further amplified by the voltage amplifier of the signal processing circuit, and adjusted to a suitable level range by the level conversion circuit of the signal processing circuit.

6. The bullet impact point positioning method according to claim 3, characterized in that: Training methods for deep learning models include: Multi-scenario live-fire data acquisition steps: Conduct a large number of live-fire experiments under different target materials, different ammunition types and different environmental noise conditions, and acquire waveform data from the three-channel piezoelectric sensor; Steps for obtaining the real time difference label: Use a high-precision assisted positioning system to record the real bullet impact point coordinates (x, y) for each shot, and combine the known coordinates of the three sensors and the propagation speed v of the shock wave in the target to back-calculate the corresponding real time difference of arrival (dt1_label, dt2_label), and use it as the supervised learning label; Data preprocessing steps: The collected three-channel waveform data are subjected to DC removal, bandpass filtering, and amplitude normalization. Waveform data are extracted according to a fixed time window, paired with supervised learning labels to form a training sample set, and divided into training set and validation set according to the ratio. Steps to build a deep learning model: Build a deep learning model using convolutional neural networks, long short-term memory networks, Transformer networks, or hybrid architectures. Take three-channel waveform data as input and output the predicted time difference of arrival (dt1) and time difference of arrival (dt2). Model training and optimization steps: Using mean squared error or mean absolute error as the loss function, measure the error between the model's predicted (dt1_pred, dt2_pred) and the true label (dt1_label, dt2_label). Use the backpropagation algorithm and Adam optimizer to iteratively update the network weights until the model converges on the validation set. Save the trained model weight file.

Images