An Enhanced High-Precision Search and Capture Method for Roland Signal Based on Convolutional Neural Networks
By employing convolutional neural network feature learning and recurrent correlation capture methods, the problem of traditional enhanced Rowland signal search and capture being susceptible to noise interference was solved, achieving high-precision signal detection and synchronization in low signal-to-noise ratio environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NAT TIME SERVICE CENT CHINESE ACAD OF SCI
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-30
AI Technical Summary
Traditional enhanced Loran signal search and capture methods are susceptible to noise and interference, leading to search failures, especially in low signal-to-noise ratio environments.
A convolutional neural network is used to learn features of the enhanced Rowland signal. Combined with the recurrent correlation capture method, the signal is searched and captured by training the neural network model, replacing the traditional sliding window cumulative search.
It significantly improves signal detection probability in low signal-to-noise ratio environments, achieves high-precision signal start-point positioning, reduces computational complexity, improves search and capture efficiency, and provides a reliable signal synchronization solution.
Smart Images
Figure CN122307608A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of land-based positioning, navigation and timing, and in particular to an enhanced Loran signal high-precision search and capture method based on convolutional neural networks. Background Technology
[0002] The Enhanced Loland System (ERS), with its low operating frequency, strong signal penetration, and strong anti-interference capabilities, has become a key backup and supplement to global navigation satellite systems. ERS utilizes pulse group signals as a carrier to establish its data communication link, enabling the remote transmission of standard time code information and other data. Different ERS stations are distinguished by different Group Repetition Periods (GRIs). The pulse groups use a two-phase, two-period phase coding system, with odd-numbered periods (GRI-A) and even-numbered periods (GRI-B), which are transmitted alternately. Taking the master station's pulse group as an example, it contains 9 pulses, and the phase coding table for the master station's GRI-A is "++--+-+-+", while the phase coding table for GRI-B is "+--+++++-". In addition, the pulses in the pulse group also employ three-state pulse-shift word modulation (PPM), which modulates information onto the 3rd to 8th pulses of the group repetition cycle through transmit timing control. The modulation amount controlled by the timing control has three states: ±1μs and 0μs. The modulation pattern on the 3rd to 8th pulses in each pulse group corresponds to a balance pattern. These balance patterns are predefined and can be found in the relevant design documents. An example of the enhanced Loran system pulse group phase coding and modulation mode can be found here. Figure 1 As shown.
[0003] Search and acquisition are the prerequisites and core components for enhanced Loran receivers to achieve signal synchronization and complete positioning and timing calculations. Their performance directly determines the receiver's ability to detect weak signals, its anti-interference capability, and its synchronization speed. Traditional search and acquisition methods mainly involve "cumulative search and correlation acquisition." This involves first performing a sliding window cumulative search on the received signal. When the accumulated signal clearly contains the enhanced Loran signal within a GRI (Gross Interval), the search is complete. Then, correlation is used to determine the starting position of the enhanced Loran signal, completing the acquisition. Because the search process requires judgment based on the characteristics of the enhanced Loran signal, its judgment threshold is highly susceptible to noise and interference, leading to search failures and affecting subsequent performance. In increasingly complex electromagnetic environments, alternative methods are needed to improve the search and acquisition capabilities of enhanced Loran systems. Summary of the Invention
[0004] The main objective of this application is to provide a high-precision search and capture method for enhanced Rowland signals based on convolutional neural networks. This method utilizes the learning ability of convolutional neural networks to learn signal characteristics, trains a suitable neural network model, applies it to the search process of enhanced Rowland signals, and combines it with a cyclic correlation capture method to complete the search and capture process under low signal-to-noise ratio conditions.
[0005] To achieve the above objectives, this application provides an enhanced high-precision search and capture method for Rowland signals based on convolutional neural networks, comprising: Obtain a target signal sequence within a group repetition period GRI range; The simulation generates the data required for neural network training. The convolutional neural network is trained using the data and the network model parameters are saved to obtain the signal search and capture model. The target signal sequence is judged using the saved signal search and capture model. If the judgment result is that it contains an enhanced Loran signal, the relevant capture process is executed. During the correlation acquisition process, a local pulse group sequence is generated. The local pulse group sequence is then cyclically correlated with the target signal sequence to obtain the maximum value of the correlation value sequence and the corresponding sampling position. Based on the maximum value of the correlation value sequence and the corresponding sampling position, the signal starting point is determined for target search and acquisition.
[0006] Optionally, the data required for training the simulated neural network includes: Enhanced Rowland signal sample data within the GRI range is generated based on the sampling rate simulation. In particular, the phase encoding, modulation pattern, and pulse group start position of the enhanced Loran signal sample data are all randomly set; Random noise of different intensities is added to the sample data of the enhanced Rowland signal to obtain a noisy enhanced Rowland signal; By superimposing a noisy enhanced Rowland signal onto a pre-constructed representation function indicating the presence or absence of an enhanced Rowland signal, input samples and output samples corresponding to each input sample are constructed from the pre-constructed representation function output, thereby generating the data required for neural network training.
[0007] Optionally, in the function representing the presence or absence of the enhanced Rowland signal, 1 indicates that the sample contains a noisy enhanced Rowland signal, and 0 indicates that the sample does not contain a noisy enhanced Rowland signal. The number of samples containing the noisy enhanced Rowland signal is approximately the same as the number of samples without the noisy enhanced Rowland signal.
[0008] Optionally, training the convolutional neural network using the data required for neural network training includes: The training data is normalized and formatted into the input format required for network training. The training data is divided into a training set and a validation set; Design the parameters of the convolutional neural network and train the convolutional neural network using the training set and validation set.
[0009] Optionally, the step of using a signal search and capture model to determine the target signal sequence includes: If the signal search and capture model outputs 0, it means that the sequence does not contain an enhanced Rowland signal. Repeat the operation of obtaining a target signal sequence within a set repetition period GRI range. If the signal search and capture model outputs 1, it indicates that the sequence contains an enhanced Rowland signal, and the relevant capture process is executed.
[0010] Optionally, the step of generating a local pulse group sequence, performing cyclic correlation calculation between the local pulse group sequence and the target signal sequence to obtain the maximum value of the correlation value sequence and the corresponding sampling position, includes: Based on the sampling rate, noise-free pulse group sequences after GRI-A and GRI-B phase encoding are generated respectively. During the generation process, the modulation pattern of the noise-free pulse group sequence is selected as a preset state, and the sequence length is consistent with the length of the acquired target signal sequence. The two noiseless pulse group sequences after generating GRI-A and GRI-B phase encoding are respectively subjected to cyclic correlation calculation with the acquired target signal sequence to obtain two correlation value sequences. Find the maximum value and the corresponding sampling position in the two correlation value sequences respectively.
[0011] Optionally, the signal search and capture model includes an input layer, a first convolutional block, a second convolutional block, a global pooling and flattening module, a fully connected and regularized module, and an output layer; The input layer is used to receive one-dimensional sequence data; The first convolutional block and the second convolutional block are used to extract local features and high-order abstract features of the sequence in sequence, respectively. The global pooling and flattening module is used to compress the feature dimension and convert it into a fixed-length feature vector; The fully connected and regularized module is used for feature mapping and preventing overfitting; The output layer is used to output the classification results.
[0012] Optionally, the first convolutional block includes a first one-dimensional convolutional layer, a first one-dimensional batch normalization layer, a first ReLU activation layer, and a first max pooling layer connected in sequence. The second convolutional block includes a second one-dimensional convolutional layer, a second one-dimensional batch normalization layer, a second ReLU activation layer, and a second max pooling layer connected in sequence.
[0013] Optionally, the global pooling and flattening module includes a global pooling layer and a flattening layer; The global pooling layer is used to compress each feature channel into a single value, eliminating the sequence length dimension; The flattening layer is used to convert multidimensional feature vectors into one-dimensional vectors.
[0014] Optionally, the fully connected and regularization module includes a first fully connected layer, a third ReLU activation layer, and a Dropout layer; The first fully connected layer is used to map global features to a higher-dimensional feature space; The Dropout layer is used to randomly deactivate some neurons to improve generalization ability.
[0015] This application proposes a high-precision search and capture method for enhanced Loran signals based on convolutional neural networks. Compared with existing technologies, this application has the following advantages: 1. By using convolutional neural networks to learn features of enhanced Loran signals, it can effectively identify signals in low signal-to-noise ratio (SNR) environments (e.g., -15dB to 0dB), overcoming the shortcomings of traditional cumulative search methods that are susceptible to noise interference, and significantly improving the signal detection probability; 2. After filtering effective signals through neural networks, combined with biphase coding cyclic correlation capture, high-precision signal start point localization is achieved. When the SNR is better than -5dB, the capture error is less than 1μs, improving the subsequent demodulation and decoding accuracy; 3. The model training process uses diverse sample data to enhance generalization ability, making it applicable to different phase coding, modulation patterns, and noisy environments, thus improving the robustness of the method; 4. It replaces the traditional sliding window cumulative search, reducing computational complexity, improving search and capture efficiency, and providing a reliable signal synchronization solution for enhanced Loran systems. Attached Figure Description
[0016] Figure 1 This is a schematic diagram illustrating an example of pulse group phase coding and modulation mode of an enhanced Loran system provided in an embodiment of this application; Figure 2 A flowchart illustrating a search and capture method based on a convolutional neural network provided in an embodiment of this application; Figure 3 This is a schematic diagram of the training progress of a convolutional neural network provided in one embodiment of this application; Figure 4 This application provides a relevant capture example and judgment diagram as an embodiment; Figure 5 This is a schematic diagram illustrating the change of search capture probability with signal-to-noise ratio according to an embodiment of this application; Figure 6 This is a schematic diagram illustrating the change of acquisition error with signal-to-noise ratio according to an embodiment of this application; Figure 7 This is a flowchart of a search and capture method based on a convolutional neural network provided in an embodiment of this application.
[0017] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0018] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.
[0019] like Figure 2 and Figure 7 As shown, the first embodiment of this application provides a high-precision search and capture method for enhanced Rowland signals based on convolutional neural networks. This embodiment uses an enhanced Rowland signal with a GRI of 60000 as an example to process a pulse group signal within a GRI range. The signal sampling rate is 1MHz. The high-precision search and capture method for enhanced Rowland signals based on convolutional neural networks may include: An enhanced high-precision search and capture method for the Rowland signal based on convolutional neural networks includes: S101. Obtain a target signal sequence within the range of a group repetition period GRI; In this process, based on the GRI value of the station to be searched and the sampling rate, a target signal sequence within a GRI range is obtained. This sequence is the signal to be searched and captured. Specifically, in this embodiment, GRI = 60000, the sampling rate is 1MHz, the number of data points within a GRI range is 60000, and the corresponding time is 0:1:59999. The unit is... μs The sequence contains an enhanced Rowland signal with a signal-to-noise ratio of 0 dB, phase-coded modulation of GRI-A, pattern modulation of "-0++-0", and a signal start position corresponding to time 43220. It is understandable that by acquiring a target signal sequence within a GRI range, a complete analytical object can be provided for subsequent neural network judgment and correlation capture, ensuring the integrity of the signal characteristics.
[0020] S102. Simulate and generate the data required for neural network training, use the data required for neural network training to train the convolutional neural network and save the network model parameters to obtain the signal search and capture model. In one embodiment of this application, the simulation to generate the data required for neural network training includes: simulating and generating enhanced Rowland signal sample data within the GRI range according to the sampling rate; wherein the phase encoding, modulation pattern, and pulse group start position of the enhanced Rowland signal sample data are all randomly set; adding random noise of different intensities to the enhanced Rowland signal sample data to obtain a noisy enhanced Rowland signal; and constructing input samples and output samples corresponding to each input sample by superimposing the noisy enhanced Rowland signal on a pre-constructed representation function of the presence or absence of the enhanced Rowland signal, so as to generate the data required for neural network training.
[0021] In the function representing the presence or absence of the enhanced Rowland signal, 1 indicates that the sample contains the noisy enhanced Rowland signal, and 0 indicates that the sample does not contain the noisy enhanced Rowland signal; the number of samples containing the noisy enhanced Rowland signal is approximately the same as the number of samples not containing the noisy enhanced Rowland signal.
[0022] In this embodiment, training the convolutional neural network using the data required for neural network training includes: normalizing the training data and organizing it into the input format required for network training; dividing the training data into a training set and a validation set; designing the parameters of the convolutional neural network; and training the convolutional neural network using the training set and the validation set.
[0023] In this embodiment, the signal search and capture model includes an input layer, a first convolutional block, a second convolutional block, a global pooling and flattening module, a fully connected and regularized module, and an output layer. The input layer is used to receive one-dimensional sequence data. The first and second convolutional blocks are used to extract local features and higher-order abstract features of the sequence sequentially. The global pooling and flattening module is used to compress the feature dimension and convert it into a fixed-length feature vector. The fully connected and regularized module is used for feature mapping and preventing overfitting. The output layer is used to output the classification result.
[0024] In an embodiment, the first convolutional block includes a first one-dimensional convolutional layer, a first one-dimensional batch normalization layer, a first ReLU activation layer, and a first max pooling layer connected in sequence; the second convolutional block includes a second one-dimensional convolutional layer, a second one-dimensional batch normalization layer, a second ReLU activation layer, and a second max pooling layer connected in sequence.
[0025] In this embodiment, the global pooling and flattening module includes a global pooling layer and a flattening layer; the global pooling layer is used to compress each feature channel into a single value, eliminating the sequence length dimension; the flattening layer is used to convert multidimensional feature vectors into one-dimensional vectors.
[0026] In this embodiment, the fully connected and regularization module includes a first fully connected layer, a third ReLU activation layer, and a Dropout layer; the first fully connected layer is used to map global features to a higher-dimensional feature space; the Dropout layer is used to randomly deactivate some neurons to improve generalization ability.
[0027] For example, enhanced Rowland signal sample data within the GRI range is generated through simulation based on the sampling rate. The signal's phase encoding, modulation pattern, and pulse group start position are all randomized. Based on the sampling rate, each sample contains 60,000 data points. The phase encoding is randomly selected from GRI-A or GRI-B. The enhanced Rowland pulse group data in the sample is randomly modulated from 128 patterns, and the pulse group start position is also randomized.
[0028] For example, random noise is superimposed on the sample data, and the signal-to-noise ratio (SNR) is selected to be between -15dB and 0dB, with an SNR interval of 1dB. 300 samples are generated in simulation at each SNR. Whether an enhanced Rowland signal is present is determined by the function randi([0,1],1), where 1 represents the presence of an enhanced Rowland signal and 0 represents the absence of an enhanced Rowland signal. The mapping label data sequence corresponding to whether a Rowland signal is present at each SNR is recorded.
[0029] Optionally, the processor can divide the training data into a training set and a validation set in a 4:1 ratio. The training samples at all signal-to-noise ratios are combined to form a 60000×3840 matrix, with the corresponding mapping label data being a 3840×1 matrix. Similarly, the validation set is a 60000×960 matrix, with the corresponding mapping label data being a 960×1 matrix. For example... Figure 3 The diagram shows the training progress of the convolutional neural network. It should be noted that when the processor uses the convolutional network to train the simulation data, it normalizes the training data, organizes it into the format required for network training, designs the convolutional network parameters, trains the network, and saves the trained network parameters for later use.
[0030] In operation, the input layer receives raw one-dimensional sequence data as the initial input to the entire network. The first convolutional block consists of a series of interconnected one-dimensional convolutional layers, a one-dimensional batch normalization layer, a ReLU activation function, and a max pooling layer, used to extract local features of the sequence. The second convolutional block further extracts more abstract, higher-order sequence features. The global pooling layer compresses each feature channel into a single value, eliminating the sequence length dimension. The flattening layer converts multi-dimensional feature vectors into one-dimensional vectors. The fully connected layer maps global features to a higher-dimensional feature space, and the Dropout layer randomly deactivates some neurons to prevent overfitting. The output layer outputs multi-class probabilities through the Softmax activation function. Clearly, by constructing a convolutional neural network model containing convolutional blocks, pooling layers, and regularization modules, and training it with randomly generated, diverse sample data, the model can deeply learn the features of the enhanced Rowland signal under low signal-to-noise ratio conditions, effectively improving the model's generalization ability and recognition accuracy. This allows it to replace traditional cumulative search during the search process, overcoming noise interference.
[0031] Specifically, the network model parameters are as follows: The first one-dimensional convolutional module contains a convolutional layer (kernel size = 1000, number of channels = 4, stride = 5, padding method "same"), batch normalization, a linear unit activation function with leakage correction (negative slope = 0.01), and a max pooling layer (pooling kernel size = 8, stride = 4) for local feature extraction and sequence downsampling. A second one-dimensional convolutional module (kernel size = 500, number of channels = 4, stride = 4, padding method "same") is stacked and uses the same normalization / activation / pooling strategy as the first module to extract high-level abstract features.
[0032] Global average pooling is used to aggregate temporal features into a fixed-length vector, which is then flattened and input into subsequent fully connected layers. The classification head consists of a fully connected layer (32 neurons, ReLU activation function), a dropout layer to mitigate overfitting (dropout rate = 0.2), and a final fully connected layer for binary classification. Softmax normalization and cross-entropy loss function are used to calculate probabilities and optimize the model.
[0033] The model is trained using the Adam optimizer (initial learning rate = 1e-3, L2 regularization coefficient = 1e-4, mini-batch size = 64) and a segmented learning rate scheduling strategy (decay factor = 0.5, decay period = 5 training epochs). Validation is performed on an independently partitioned validation set, with an early stopping patience value set to 50 training epochs; to avoid order bias, the training data is randomly shuffled in each training epoch.
[0034] S103. Use the saved signal search and capture model to judge the target signal sequence. If the judgment result is that it contains an enhanced Loran signal, then execute the relevant capture process. In one embodiment of this application, the step of using a signal search and capture model to determine the target signal sequence includes: if the output result of the signal search and capture model is 0, it indicates that the sequence does not contain an enhanced Rowland signal, and the operation of obtaining a target signal sequence within a set repetition period (GRI) range is repeated; if the output result of the signal search and capture model is 1, it indicates that the sequence contains an enhanced Rowland signal, and the relevant capture process is executed.
[0035] In this process, the saved network model structure is used to determine whether the target signal sequence to be judged in step S101 contains an enhanced Rowland signal. Specifically, in this example, the output result is 1, indicating that the sequence contains an enhanced Rowland signal, and the correlation capture process continues. If the output result is 0, it indicates that the sequence does not contain an enhanced Rowland signal, and the operation of step S101 is repeated to enter the next loop. To be precise, through the pre-judgment of the neural network model, invalid data segments that do not contain the signal can be quickly eliminated, avoiding complex correlation operations on invalid data, thereby significantly reducing the amount of computation and improving the efficiency of search and capture.
[0036] S104. During the correlation acquisition process, a local pulse group sequence is generated. The local pulse group sequence is cyclically correlated with the target signal sequence to obtain the maximum value of the correlation value sequence and the corresponding sampling position. The signal starting point is determined based on the maximum value of the correlation value sequence and the corresponding sampling position to search for and acquire the target.
[0037] In one embodiment of this application, the step of generating a local pulse group sequence, which involves performing cyclic correlation calculations between the local pulse group sequence and the target signal sequence to obtain the maximum value and corresponding sampling position of the correlation value sequence, includes: generating noise-free pulse group sequences after GRI-A and GRI-B phase encoding respectively according to the sampling rate, and during the generation process, selecting the modulation pattern of the noise-free pulse group sequence as a preset state, with the sequence length consistent with the length of the acquired target signal sequence; performing cyclic correlation calculations between the generated GRI-A and GRI-B phase-encoded noise-free pulse group sequences and the acquired target signal sequence respectively to obtain two correlation value sequences; and finding the maximum value and corresponding sampling position of the two correlation value sequences respectively.
[0038] Specifically, when the prediction result in step S103 is 1, noise-free pulse group sequences after GRI-A and GRI-B phase encoding are generated according to the sampling rate, with the modulation pattern selected as "000000", and the sequence length is consistent with the sequence length in step S101. Specifically, cyclic correlation calculations are performed between these sequences and the sequences in step S101 to obtain two correlation value sequences. The maximum value and corresponding sampling position in each correlation value sequence are then identified. For example,... Figure 4 As shown, in this example, the maximum value associated with GRI-A is 341.3018, corresponding to a time of 43220, while the maximum value associated with GRI-B is 125.0016, corresponding to a time of 37220. Comparing the maximum values in the two correlation sequences, 341.3018 > 125.0016. The sampling position corresponding to the larger of the two maximum values is the starting point of the signal, indicating that the signal has been captured and is a phase-coded GRI-A signal. In fact, after filtering out effective signals through a neural network, and then combining it with a biphase-coded cyclic correlation capture method, the starting point of the signal can be accurately located even at a low signal-to-noise ratio. This overcomes the problem of search failure caused by noise interference in traditional methods at low signal-to-noise ratios, achieving high-precision signal capture.
[0039] As can be seen from the above embodiments, the high-precision search and capture method for enhanced Rowland signals based on convolutional neural networks can achieve the search and capture of enhanced Rowland signals. The probability of search and capture can be statistically calculated from the results of multiple trials at different signal-to-noise ratios (2000 samples per signal-to-noise ratio). The probability value represents the proportion of sequences correctly identified by the capture correlation method in the enhanced Rowland signal out of the total number of sequences. Compared with traditional search and capture, the statistical probabilities at different signal-to-noise ratios are as follows: Figure 5 As shown, this method significantly outperforms the cumulative correlation capture method, achieving a capture probability of approximately 65% based on convolutional neural network search at a signal-to-noise ratio of -10dB. Figure 6 The figure shows the relationship between the acquisition error and the signal-to-noise ratio (SNR) after acquisition using a convolutional neural network (CNN) method followed by sliding correlation acquisition. It can be seen from the figure that when the SNR is better than -5dB, the acquisition error is less than 1. The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
Claims
1. A high-precision search and capture method for enhanced Rowland signals based on convolutional neural networks, characterized in that, include: Obtain a target signal sequence within a group repetition period GRI range; The simulation generates the data required for neural network training. The convolutional neural network is trained using the data and the network model parameters are saved to obtain the signal search and capture model. The target signal sequence is judged using the saved signal search and capture model. If the judgment result is that it contains an enhanced Loran signal, the relevant capture process is executed. During the correlation acquisition process, a local pulse group sequence is generated. The local pulse group sequence is then cyclically correlated with the target signal sequence to obtain the maximum value of the correlation value sequence and the corresponding sampling position. Based on the maximum value of the correlation value sequence and the corresponding sampling position, the signal starting point is determined for target search and acquisition.
2. The enhanced Roland signal high-precision search and capture method based on convolutional neural networks as described in claim 1, characterized in that, The data required for training the simulated generative neural network includes: Enhanced Rowland signal sample data within the GRI range is generated based on the sampling rate simulation. In particular, the phase encoding, modulation pattern, and pulse group start position of the enhanced Loran signal sample data are all randomly set; Random noise of different intensities is added to the sample data of the enhanced Rowland signal to obtain a noisy enhanced Rowland signal; By superimposing a noisy enhanced Rowland signal onto a pre-constructed representation function indicating the presence or absence of an enhanced Rowland signal, input samples and output samples corresponding to each input sample are constructed from the pre-constructed representation function output, thereby generating the data required for neural network training.
3. The enhanced Roland signal high-precision search and capture method based on convolutional neural networks as described in claim 2, characterized in that, In the function representing the presence or absence of the enhanced Rowland signal, 1 indicates that the sample contains a noisy enhanced Rowland signal, and 0 indicates that the sample does not contain a noisy enhanced Rowland signal. The number of samples containing the noisy enhanced Rowland signal is approximately the same as the number of samples without the noisy enhanced Rowland signal.
4. The enhanced Roland signal high-precision search and capture method based on convolutional neural networks as described in claim 1, characterized in that, The process of training a convolutional neural network using the data required for neural network training includes: The training data is normalized and formatted into the input format required for network training. The training data is divided into a training set and a validation set; Design the parameters of the convolutional neural network and train the convolutional neural network using the training set and validation set.
5. The enhanced Roland signal high-precision search and capture method based on convolutional neural networks as described in claim 2, characterized in that, The process of using a signal search and capture model to determine the target signal sequence includes: If the signal search and capture model outputs 0, it means that the sequence does not contain an enhanced Rowland signal. Repeat the operation of obtaining a target signal sequence within a set repetition period GRI range. If the signal search and capture model outputs 1, it indicates that the sequence contains an enhanced Rowland signal, and the relevant capture process is executed.
6. The enhanced Roland signal high-precision search and capture method based on convolutional neural networks as described in claim 1, characterized in that, The process of generating a local pulse group sequence involves performing a cyclic correlation calculation between the local pulse group sequence and the target signal sequence to obtain the maximum value of the correlation value sequence and the corresponding sampling position, including: Based on the sampling rate, noise-free pulse group sequences after GRI-A and GRI-B phase encoding are generated respectively. During the generation process, the modulation pattern of the noise-free pulse group sequence is selected as a preset state, and the sequence length is consistent with the length of the acquired target signal sequence. The two noiseless pulse group sequences after generating GRI-A and GRI-B phase encoding are respectively subjected to cyclic correlation calculation with the acquired target signal sequence to obtain two correlation value sequences. Find the maximum value and the corresponding sampling position in the two correlation value sequences respectively.
7. The enhanced Roland signal high-precision search and capture method based on convolutional neural networks as described in claim 1, characterized in that, The signal search and capture model includes an input layer, a first convolutional block, a second convolutional block, a global pooling and flattening module, a fully connected and regularized module, and an output layer. The input layer is used to receive one-dimensional sequence data; The first convolutional block and the second convolutional block are used to extract local features and high-order abstract features of the sequence in sequence, respectively. The global pooling and flattening module is used to compress the feature dimension and convert it into a fixed-length feature vector; The fully connected and regularized module is used for feature mapping and preventing overfitting; The output layer is used to output the classification results.
8. The enhanced Roland signal high-precision search and capture method based on convolutional neural networks as described in claim 7, characterized in that, The first convolutional block includes a first one-dimensional convolutional layer, a first one-dimensional batch normalization layer, a first ReLU activation layer, and a first max pooling layer connected in sequence. The second convolutional block includes a second one-dimensional convolutional layer, a second one-dimensional batch normalization layer, a second ReLU activation layer, and a second max pooling layer connected in sequence.
9. The enhanced Roland signal high-precision search and capture method based on convolutional neural networks as described in claim 7, characterized in that, The global pooling and flattening module includes a global pooling layer and a flattening layer; The global pooling layer is used to compress each feature channel into a single value, eliminating the sequence length dimension; The flattening layer is used to convert multidimensional feature vectors into one-dimensional vectors.
10. The enhanced Roland signal high-precision search and capture method based on convolutional neural networks as described in claim 7, characterized in that, The fully connected and regularization module includes a first fully connected layer, a third ReLU activation layer, and a Dropout layer; The first fully connected layer is used to map global features to a higher-dimensional feature space; The Dropout layer is used to randomly deactivate some neurons to improve generalization ability.