Strong coherence multipath time delay estimation method based on deep learning
By training the method of delaying the multi -way period based on deep learning, and using sparse encoding and multi -heading mechanism to train deep convolutional networks to solve the problem of lack of ability to distinguish the ability to distinguish the ability in the scene during the strong scattered multi -way, Implement high -precision delay estimates.
Patent Information
- Application Number
- CN202411501086.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-25
- Publication Date
- 2025-01-28
AI Technical Summary
The existing methods lack the resolution capability in strong coherent multi-path delay estimation scenarios, especially when the signal interval is one-part of the bandwidth, resulting in poor multi-path delay estimation accuracy.
Adopting a strong -way multi -way extension method based on deep learning, training deep convolutional networks through sparse coding and multi -header attention mechanism to achieve high -precision delay estimates.
During the extension of the extension of the extension, the signal recognition ability and estimation accuracy were significantly improved, and the accuracy of the signal recognition was significantly improved.
Smart Images
Figure CN119382804A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of underwater acoustic signal detection, and in particular to a strong coherence multipath delay estimation method based on deep learning. Background Art
[0002] Near-field strong coherence multipath delay estimation is a research topic with broad application value in underwater positioning. Accurate multipath delay estimation is an important basis for target distance solution and signal detection. Strong coherence means that the signal similarity is high and the time domain interval is less than one-tenth of the bandwidth.
[0003] When using ultra-short baselines to locate targets, the receiver receives signals sent simultaneously by multiple elements of the transmitter and signals reflected by adjacent elements or array mounting wall panels. In this scenario, the receiver faces the problem of multipath interference when estimating the delay of each signal. Existing methods include: matched filtering algorithm based on correlation, super-resolution algorithm based on subspace decomposition, and delay estimation algorithm based on deep learning.
[0004] In the process of implementing the embodiments of the present invention, the inventors found that the background technology has the following defects:
[0005] The correlation-based matched filtering algorithm has no ability to resolve aliased multipath signals within a spacing of one-fifth of the bandwidth.
[0006] The performance of delay estimation algorithms based on subspace decomposition, such as the Multiple Signal Classification (MUSIC) algorithm, drops sharply in complex noise interference scenarios because it is based on the principle of spatial orthogonality between signal and noise. In addition, the algorithm has poor ability to distinguish aliased multipath signals with an interval of one-third of the bandwidth and cannot meet application requirements.
[0007] The problem with using deep learning algorithms based on feature extraction and training the abstract functional relationship between features and labels is that there is a gap in the application of time delay estimation in strongly coherent multi-path scenarios. In addition, there is little research on algorithm model optimization in the field of underwater acoustic signal delay estimation, and there is room for optimization in estimation accuracy and generalization. Summary of the invention
[0008] The purpose of the present invention is to solve the problem that the existing methods cannot meet the requirements for resolving aliased multipath signals within an interval of one-third of the bandwidth, resulting in poor accuracy of multipath delay estimation, and to propose a strongly coherent multipath delay estimation method based on deep learning.
[0009] A near-field strong coherent multipath delay estimation method based on deep learning algorithm. The specific process is as follows:
[0010] Step 1: obtain a data packet by simulation, the data packet including signal data received by the array element and a delay true value label corresponding to the signal data;
[0011] Step 2: The signal data in the data packet is sequentially converted into data type, outliers are eliminated, the real and imaginary parts of the data are modulo-ed, and the data is normalized to obtain the processed signal data;
[0012] Step 3: Sparsely encode the true value of the delay corresponding to the signal data in the data packet using a sparse coding method based on the Sinc kernel function to obtain the delay value after sparse processing;
[0013] Step 4: Use the signal data processed in step 2 as the input of the delay regression network structure model, use the sparsely processed delay value obtained in step 3 as the output of the delay regression network structure model, train the delay regression module network structure model until convergence, and obtain a trained delay regression module network structure model;
[0014] Step 5: Process the signal data in the data packet of step 1 to obtain processed signal data;
[0015] The processed signal data is input into the trained delay regression module network structure model, and the trained delay regression module network structure model outputs the value corresponding to each processed signal in the preset delay grid;
[0016] The peak-finding function is used to process the value corresponding to the preset delay grid of each processed signal output by the trained delay regression module network structure model to obtain the value corresponding to the original delay grid.
[0017] The beneficial effects of the present invention are:
[0018] The inventive concept of the present invention is to introduce a deep convolutional network based on sparse coding and a multi-head attention mechanism. In the delay estimation process, the data is first sparsely encoded and used as a training label, and then the deep convolutional network is used to train the abstract functional relationship between the input data and the sparsely processed label, thereby achieving the purpose of multi-path delay estimation.
[0019] The algorithm proposed in the present invention has a strong signal recognition ability in the scenario of strong coherence multipath delay estimation, can break through the constraint of delay estimation resolution, accurately estimate the delay of multipath signals with a sampling point interval of one-third of the bandwidth, improves the multipath delay estimation accuracy, and has a relatively stable delay estimation performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 This is the overall algorithm flow chart;
[0021] Figure 2 Flowchart for sparse coding calculation;
[0022] Figure 3 This is the network structure diagram of the time delay regression module;
[0023] Figure 4 The change graph of training set loss when the value is 20dB;
[0024] Figure 5 Scatter plot of delay estimation error between the existing deep learning algorithm and the algorithm of the present invention when the delay is 20dB;
[0025] Figure 6 This is a comparison diagram of the mean square error curve of the delay estimation obtained by the ablation experiment under different signal-to-noise ratios of the present invention;
[0026] Figure 7 It is a schematic diagram showing that each simulation signal data of the present invention is obtained by aliasing two signals;
[0027] Figure 8 It is a delay grid diagram. DETAILED DESCRIPTION
[0028] Specific implementation method 1: This implementation method is a near-field strong coherent multipath delay estimation method based on a deep learning algorithm. The specific process is as follows:
[0029] Step 1: obtain a data packet by simulation, the data packet including signal data received by the array element and a delay true value label corresponding to the signal data;
[0030] Taking the two-path signal as an example, each simulated signal data is obtained by mixing two signals, and thus corresponds to two labels, such as Figure 7 ;
[0031] Assume that signal 1 (red) has a real delay value of 1×10 -5 s, the corresponding delay grid is 1
[0032] Assume that the actual delay value of signal 2 (blue) is 4×10 -5 s, the corresponding delay grid is 4
[0033] The two signals are superimposed to obtain a simulated signal. The label in the data packet is (1×10 -5 ,4×10 -5 ), the corresponding grid value is (1,4);
[0034] The delay grid is obtained as Figure 8 As shown;
[0035] Step 2: The signal data in the data packet is processed in turn by data type conversion, outlier removal, modulus of real and imaginary parts of the data, and data normalization to obtain processed signal data;
[0036] Step 2 only processes the data in the data packet to obtain processed data;
[0037] Step 3: Only the true delay value label corresponding to the data in the data packet is sparsely processed to obtain the delay value after sparse processing, and the delay value after sparse processing is the label of the data in the data packet;
[0038] Step 3: Sparsely encode the true value of the delay corresponding to the signal data in the data packet using a sparse coding method based on the Sinc kernel function to obtain the delay value after sparse processing;
[0039] Step 4: Use the signal data processed in step 2 as the input of the delay regression network structure model, use the sparsely processed delay value obtained in step 3 as the output of the delay regression network structure model, train the delay regression module network structure model until convergence, and obtain a trained delay regression module network structure model;
[0040] Step 5: Process the signal data (only data, without corresponding labels) in the data packet of step 1 to obtain processed signal data;
[0041] The processed signal data is input into the trained delay regression module network structure model, and the trained delay regression module network structure model outputs the value corresponding to each processed signal in the preset delay grid;
[0042] The peak-finding function is used to process each processed signal (200 data, each data has multiple delay labels, and each data has multiple peaks) output by the trained delay regression module network structure model at the value corresponding to the preset delay grid to obtain the value corresponding to the original delay grid.
[0043] The present invention can break through the constraint of delay estimation resolution and accurately estimate the delay of a multipath signal with a sampling point interval of one-third of the bandwidth.
[0044] Specific implementation method 2: This implementation method is different from specific implementation method 1 in that in step 2, the signal data in the data packet is sequentially processed by data type conversion, outlier removal, modulus of real and imaginary parts of the data, data normalization, etc. to obtain the processed signal data;
[0045] The specific process is:
[0046] Step 21: Convert the data type of the signal data in the data packet to obtain Numpy type data;
[0047] Step 22: Remove outliers from the obtained Numpy data (outliers refer to missing data);
[0048] Step 23: Take the modulus value of the real part and the imaginary part of the Numpy type data after removing the abnormal values, and obtain the data after the modulus value;
[0049] Step 24: normalize the data after taking the modulus value to obtain the processed signal data.
[0050] The other steps and parameters are the same as those in the first embodiment.
[0051] Specific implementation method three: This implementation method is different from specific implementation method one or two in that in step three, the delay true value corresponding to the signal data in the data packet is sparsely processed by a sparse coding method based on a Sinc kernel function to obtain a sparsely processed delay value;
[0052] The specific process is:
[0053] The present invention uses the Sinc function to process the function value corresponding to the delay grid to obtain a code with more sidelobe information. The steps are as follows: Figure 2 As shown;
[0054] Step 31: Set the preset delay grid size and interval;
[0055] Delay grid horizontal axis: The preset delay grid range is 0-20, and the corresponding real time domain range is 0s-20×10 - 5 s, the delay grid with a range of 0-20 is divided into 1000 small grids (that is, the grid resolution (interval between adjacent grids) is set to 0.02), and the real time domain corresponding to each small grid is 20×10 -8 s;
[0056] 1000 small grids means that a delay vector is composed of 1000 0s or 1s;
[0057] Delay grid ordinate: delay value after sparse processing;
[0058] Step 32: Find the delay grid closest to the true delay value corresponding to the signal data in the data packet, and thin out the delay value (such as 0.1) corresponding to the delay grid closest to the true delay value (such as 0.11) into a set of vectors consisting of 0 and 1;
[0059] Step 3: Use Sinc to perform convolution calculation on the vector composed of 0 and 1 to obtain the convolution result p1;
[0060] Step 34: perform weighted calculation on the actual delay value (such as 0.11) and the out-of-grid portion (such as 0.11-0.1=0.01) of the nearest preset grid value (such as 0.1) to obtain a weighted calculation result p2;
[0061] Step 35: Add p1 and p2 to obtain the delay value after sparse processing.
[0062] The delay value after sparse processing is the label of the data in the data packet;
[0063] Sparse coding is used to represent the true values of TOA parameters, and the network is optimized by minimizing the error between the output of the trained network and the sparse representation of the trained TOA.
[0064] The sparse coding module is based on the Sinc kernel function: after processing the real delay value corresponding to the delay grid using the Sinc function, a sparse coding with more sidelobe information can be obtained, so that the present invention has more superior accuracy and stability.
[0065] The other steps and parameters are the same as those in the first or second embodiment.
[0066] Specific implementation method 4: This implementation method is different from specific implementation methods 1 to 3 in that, in step 32, the delay grid closest to the true delay value corresponding to the signal data in the data packet is found, and the delay value (such as 0.1) corresponding to the delay grid closest to the true delay value (such as 0.11) is sparsely converted into a set of vectors consisting of 0 and 1; the specific process is:
[0067]
[0068] in, is the preset delay grid value, δ() is the unit pulse function, D is the total number of signal paths, and i is the number of the i-th signal path.
[0069] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.
[0070] Specific implementation mode 5: This implementation mode is different from the specific implementation modes 1 to 4 in that in step 33, Sinc is used to perform convolution calculation on the vector composed of 0 and 1 to obtain the convolution result p1;
[0071] The calculation formula is as follows:
[0072]
[0073] in, The delay grid value closest to the actual delay value; is the convolution calculation, and sinc() is the sinc function.
[0074] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.
[0075] Specific implementation method 6: This implementation method is different from specific implementation methods 1 to 5 in that in steps 3 and 4, the actual delay value (such as 0.11) and the out-of-grid part (such as 0.11-0.1=0.01) of the nearest preset grid value (such as 0.1) are weighted calculated to obtain the weighted calculation result p2;
[0076] The calculation formula is as follows:
[0077]
[0078] Among them, c is the set weighting coefficient, τ i It is the actual delay value of each path.
[0079] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.
[0080] Specific implementation method 7: This implementation method is different from the specific implementation methods 1 to 6 in that p1 and p2 are added in step 35 to obtain the delay value after sparse processing, and the formula is as follows:
[0081] p=p1+p2
[0082] Where p is the delay value after sparse processing.
[0083] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.
[0084] Specific implementation eight: This implementation differs from any one of specific implementations one to seven in that, in step four, the data processed in step two is used as the input of the delay regression network structure model, and the sparsely processed delay value obtained in step three (the label of the data processed in step two) is used as the output of the delay regression network structure model, and the delay regression module network structure model is trained until convergence, thereby obtaining a trained delay regression module network structure model;
[0085] The specific process is:
[0086] The network structure model of the time delay regression module includes the first fully connected layer FC, the first module, the second module, the third module, the fourth module, the fifth module, the sixth module, the seventh module, the eighth module, and the deep convolution layer;
[0087] The working process of the network structure model of the time delay regression module is as follows:
[0088] The data processed in step 1 is input into the first fully connected layer FC, and the first fully connected layer FC outputs feature A;
[0089] Feature A is sequentially input into the first module, the second module, the third module, the fourth module, the fifth module, the sixth module, the seventh module, the eighth module, and the deep convolution layer, and the deep convolution layer outputs the feature;
[0090] The output feature of the deep convolutional layer is the delay value after sparse processing obtained in step 3.
[0091] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.
[0092] Specific implementation nine: This implementation is different from any of specific implementations one to eight in that each of the first module, the second module, the third module, the fourth module, the fifth module, the sixth module, the seventh module, and the eighth module includes: a first convolutional layer, a first batch normalization BN, a first ReLuctant activation function layer, a second convolutional layer, a second batch normalization BN, a second ReLuctant activation function layer, a maximum pooling layer, a second fully connected layer FC, a third ReLuctant activation function layer, a third fully connected layer FC, a fourth ReLuctant activation function layer, and a multi-head attention mechanism layer.
[0093] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.
[0094] Specific implementation method ten: This implementation method is different from the specific implementation methods one to nine in that the working process of each module in the first module, the second module, the third module, the fourth module, the fifth module, the sixth module, the seventh module, and the eighth module is as follows:
[0095] Feature A is sequentially input into the first convolutional layer, the first batch normalization BN, the first ReLuctant activation function layer, the second convolutional layer, the second batch normalization BN, the second ReLuctant activation function layer, and the second ReLuctant activation function layer outputs feature B;
[0096] Feature B is sequentially input into the maximum pooling layer, the second fully connected layer FC, the third ReLuctant activation function layer, the third fully connected layer FC, the fourth ReLuctant activation function layer, and the fourth ReLuctant activation function layer outputs feature C;
[0097] Feature B is input into the multi-head attention mechanism layer, and the multi-head attention mechanism layer outputs feature D;
[0098] Multiply feature C and feature D element by element to obtain feature E;
[0099] Feature E is added to feature A element by element to obtain feature F;
[0100] Feature F is the output feature of each module.
[0101] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.
[0102] The time delay regression is based on a deep convolutional network structure, which includes the following four parts:
[0103] 1) Convolution operation, by sliding a small filter on the input data, calculating the dot product of the filter and the local area of the input data, and obtaining the output feature map. This operation can extract the features of the input data and is a key step for feature extraction in the CNN model.
[0104] 2) Activation functions are used to introduce nonlinearity so that the model can learn more complex features. The role of these functions is to enhance the expressiveness of the model so that it can handle a wider range of input data.
[0105] 3) Pooling layer, which is used to reduce the dimension of feature maps and the amount of computation while retaining important feature information. Commonly used pooling methods include maximum pooling and average pooling. The role of the pooling layer is to further extract and concentrate features, making the model more robust to small changes in the input data.
[0106] 4) Fully connected layer, which integrates the features extracted by the previous convolutional layer and pooling layer, performs linear transformation through weights and biases, and then performs nonlinear transformation through activation functions to finally obtain the classification result. The function of the fully connected layer is to map the learned features to the sample label space to complete the classification or regression task.
[0107] The time-delayed regression is based on a multi-head attention mechanism: The multi-head attention mechanism can capture the connections and differences in the sequence in more detail by computing and merging attention scores in parallel. This mechanism processes the input data by dividing it into multiple "heads", each of which processes the input data independently, thereby better expressing the meaning and association of each word. The split operation of the multi-head attention allows different heads to process the input data independently, while the merge operation merges the attention scores of each head into one score to generate the final attention representation.
[0108] The present invention is based on a multi-head attention mechanism: the cleaned data is processed as follows:
[0109] 1) Splitting operation: The input data is divided into multiple parts, and each head processes the input data independently, so that the attention scores can be calculated in parallel and the computational efficiency can be improved.
[0110] 2) Calculate attention scores: Each head calculates the similarity scores between its corresponding query, key, and value. These scores are used to measure the relevance between words in the sequence.
[0111] 3) Merge operation: Merge the attention scores of each head to generate the final attention representation. This step is usually done by simple summation or other aggregation methods.
[0112] The time-delay regression introduces a residual structure: In order to reduce the risks of gradient vanishing and overfitting during the training of deep neural networks, a residual structure is introduced. The residual structure enables the network to better capture and process deeper features by introducing jump connections. At the same time, by adding an input to the function output, the relationship between the input and output can still be described. This operation can be clearly split into two linear superpositions, making learning easier, while alleviating the training problem of deep networks and providing the possibility for training networks with thousands of layers.
[0113] The present invention introduces a residual structure into the time-delay regression: by introducing the residual structure, the network can better capture and process deeper features. The algorithm of the present invention is composed of a multi-layer neural network, and the gradient may gradually disappear during the back propagation process, making the network difficult to train. In order to solve the problem of gradient disappearance during training, the input is directly connected to the output so that the gradient can flow directly back from the output to the input layer, so that the purpose of effective training can be achieved while increasing the depth of the neural network.
[0114] The time-delay regression encapsulates a deep convolutional network with a residual structure into algorithm modules: in each algorithm module, two convolutional layers, one pooling layer, one attention mechanism, two fully connected layers and one residual combination. The weight of each feature is learned using a processing layer consisting of a convolutional layer, a pooling layer and a fully connected layer, and a normalization layer is added to the residual block to enhance the stability of the features in the network. These features are then input into two fully connected layers, and the multi-head attention mechanism output data is multiplied by the fully connected layer output weight. Finally, a residual structure is used to avoid features that disappear during transmission between deep layers.
[0115] In the present invention, the time delay regression encapsulates the deep convolutional network with residual structure into an algorithm module: each algorithm module consists of two convolutional layers, a pooling layer, an attention mechanism, two fully connected layers and a residual structure. The specific structure is as follows:
[0116] 1) Use two convolutional layers to learn the weight of each feature channel and extract TOA features.
[0117] 2) Use a maximum pooling layer to compress the features to obtain the global receptive field.
[0118] 3) Two fully connected layers are used to calculate the nonlinear relationship of the compressed feature channels, where the Relu function is used as the activation function, and the output data is expanded to the same dimension as the input data.
[0119] 4) The input signal is passed through a multi-head attention mechanism, the attention scores are calculated and merged in parallel, and the output is multiplied with the output weight of the fully connected layer.
[0120] 5) Use residual blocks to avoid the disappearance of part of TOA features during transmission between deep layers.
[0121] The above structure is encapsulated as an algorithm module. When presetting parameters, the number of neural network layers can be quickly increased by directly changing the number of algorithm modules. At the same time, the residual structure can effectively avoid the occurrence of gradient vanishing phenomenon and improve the performance of network feature extraction.
[0122] Example
[0123] Step 1), in the scenario of near-field strong coherent delay estimation, the present invention is simulated, and the simulation signal parameters are as follows: signal length 20ms, frequency domain range 20kHz-30kHz, sampling frequency 100kHz, number of array elements 50, signal delay value 0s to 20×10 -5 For any value in s, the multipath number is 3.
[0124] Step 2) Set the required parameters in the algorithm module. Collect 200,000 samples in total, randomly divide them into 99% for training, the batch size of each training batch is 200, and 1% is used for verification. The background noise is a white Gaussian noise corresponding sampling signal with additive noise. The training is based on Pytorch and Adam optimizer. The true TOA values of the training set and the test set are randomly generated within the specified interval. For algorithms containing sparse coding, the length of the predefined overcomplete dictionary is 1000. Calculate the root mean square error (RMSE) of all TOA estimates to evaluate the estimation accuracy.
[0125] Step 3) Use data cleaning to perform the following processing, and implement the steps as follows Figure 2 As shown, the details are as follows:
[0126] Step 3-1), perform type conversion on the array element receiving data and the active sonar transmitting signal data to obtain Numpy type data;
[0127] Step 3-2), encapsulate the data into a two-dimensional array with a dimension of 200×50 for each batch of training data;
[0128] Step 3-3), remove outliers from the converted type data to eliminate the interference of outliers on signal processing;
[0129] Step 3-4), take the modulus value of the real part and the imaginary part of the processed data and perform normalization processing to obtain standard data.
[0130] Step 4) Calculate the rough estimate of the delay using the delay rough estimation module. Figure 2 As shown, the details are as follows:
[0131] Step 4-1) Set the preset delay grid size and interval. In this simulation, the delay grid length is set to 0-20 and the grid interval is 0.02;
[0132] Step 4-2), find the delay grid closest to the actual delay value, and thin the delay value into a set of vectors consisting of 0 and 1;
[0133] Step 4-3) Perform convolution calculation on the vector composed of 0 and 1 using the Sinc function. The convolution Sinc function formula is as follows:
[0134]
[0135] Step 4-4), perform weighted calculation on the actual delay value and the out-of-grid portion of the nearest preset grid value, and set the additional value calculation weight to 100;
[0136] Step 4-5), the sparse delay value is added to the data obtained by weighted calculation to obtain a sparse expression of the final off-grid delay value.
[0137] Step 6), input the sparsely coded labels and the signal data output by the delay rough estimation module into the delay regression module network structure for training, as shown in the steps Figure 3 As shown, the details are as follows:
[0138] Step 6-1) Use the internal linear calculation of the fully connected layer FC to expand the data dimension to batchsize×4000
[0139] Step 6-2) Use two convolutional layers to learn the weight of each feature channel and extract the TOA feature.
[0140] Step 6-3) Use a maximum pooling layer to compress the features to obtain the global receptive field.
[0141] Step 6-4) Use two fully connected layers to calculate the nonlinear relationship of the compressed feature channel, where the activation function uses the Relu function, and then expand the output data to the same dimension as the input data.
[0142] Step 6-5) Pass the data obtained in step 6-1) through a multi-head attention mechanism. The multi-head attention mechanism captures the connections and differences in the sequence more carefully by parallel calculation and merging attention scores, and then multiplies the output with the output weight of the fourth ReLuctant activation function layer in each module.
[0143] Step 6-6) Use the residual block to avoid the disappearance of part of the TOA features during transmission between deep layers.
[0144] When presetting parameters, the number of modules is set to 8, with a total of 16 convolutional layers, 8 pooling layers, 16 fully connected layers, 8 multi-head attention mechanism processing and 8 residual structures. The hyperparameters of the convolution kernel size, step size and fully connected layer in the present invention are as follows: Figure 3 as shown in .
[0145] Step 7) Randomly set the weights and bias values of the input layer, and establish a nonlinear relationship between layers during the forward propagation process. During the forward propagation process, the input data passes through the input layer, the hidden layer, and is processed layer by layer and transmitted to the output layer.
[0146] Step 8) Calculate the sum of squares of the output layer and the set label error value as the constraint function, use the partial derivative of the objective function to calculate the weight of each neuron in the back propagation, and use the stochastic gradient descent algorithm to gradually correct the weights.
[0147] Step 9), repeat step 6) to step 8) until the number of training times reaches a preset value.
[0148] Step 10) Figure 4 This is a graph showing the change in the neural network training set loss when the signal-to-noise ratio is set to 20dB.
[0149] Step 11), randomly extract data to perform 120 delay estimations, and obtain the scatter plot of the delay estimation error of the present invention at 20 dB as shown in the figure: Figure 5 shown.
[0150] Step 12), repeat steps 1) to 10) at multiple signal-to-noise ratios in the range of -10 to 20 dB with a step size of 5 dB, and perform ablation experiments. The existing deep learning algorithm, the residual structure algorithm, and the residual structure and sparse coding algorithm are successively performed to obtain the mean square error curve of the delay estimation of the ablation experiment of the present invention under different signal-to-noise ratios. Figure 6 shown.
[0151] Step 13), by Figure 6 It can be seen that when the signal-to-noise ratio is -10 to 20dB and the step size is 5dB, the RMSE value of the proposed delay estimation algorithm is significantly reduced compared with the existing deep learning algorithm. That is, in the near-field strong coherent multipath delay estimation scenario, the delay estimation accuracy of the algorithm proposed in this paper is significantly improved compared with the existing deep learning delay estimation algorithm. When SNR<20dB, the algorithm RMSE value drops by more than 11%, and the decrease gradually increases with the increase of the signal-to-noise ratio. When SNR=20dB, the algorithm RMSE value drops by about 86%.
[0152] Step 14), in the delay estimation data obtained by the ablation experiment when the signal-to-noise ratio is 20 dB, the delay estimation accuracy under different confidence intervals of the existing deep learning algorithm, the residual structure addition algorithm, and the residual structure addition and sparse coding algorithm are summarized as shown in Table 1.
[0153] Step 15), as shown in Table 1, when the signal-to-noise ratio is 20dB, the RMSE value of the proposed delay estimation algorithm is significantly reduced compared with the existing deep learning algorithm. When the confidence interval is 1 / 10 of the bandwidth and larger, the accuracy of the optimized algorithm reaches 100%, which is more than 34.9% higher than the accuracy of the existing deep learning algorithm; when the confidence interval is 1 / 50 and 1 / 100 of the bandwidth, the accuracy of the proposed algorithm is 82.4% and 88.5% higher than that of the existing deep learning algorithm, respectively.
[0154] Table 1 Delay estimation accuracy obtained from ablation experiments with different confidence intervals at 20 dB
[0155]
[0156] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.
Claims
1. A method for estimating strongly coherent multipath delay based on deep learning, characterized in that: The specific process of the method is: Step 1: obtain a data packet by simulation, the data packet including signal data received by the array element and a delay true value label corresponding to the signal data; Step 2: The signal data in the data packet is sequentially converted into data type, outliers are eliminated, the real and imaginary parts of the data are modulo-ed, and the data is normalized to obtain the processed signal data; Step 3: Sparsely encode the true value of the delay corresponding to the signal data in the data packet using a sparse coding method based on the Sinc kernel function to obtain the delay value after sparse processing; Step 4: Use the signal data processed in step 2 as the input of the delay regression network structure model, use the sparsely processed delay value obtained in step 3 as the output of the delay regression network structure model, train the delay regression module network structure model until convergence, and obtain a trained delay regression module network structure model; Step 5: Process the signal data in the data packet of step 1 to obtain processed signal data; The processed signal data is input into the trained delay regression module network structure model, and the trained delay regression module network structure model outputs the value corresponding to each processed signal in the preset delay grid; The peak-finding function is used to process the value corresponding to the preset delay grid of each processed signal output by the trained delay regression module network structure model to obtain the value corresponding to the original delay grid.
2. The method for estimating strongly coherent multipath delay based on deep learning according to claim 1, characterized in that: In the step 2, the signal data in the data packet is sequentially converted into a data type, outliers are eliminated, the real and imaginary parts of the data are modulo-ed, and the data is normalized to obtain the processed signal data; The specific process is: Step 21: Convert the data type of the signal data in the data packet to obtain Numpy type data; Step 22: Remove outliers from the obtained Numpy data; Step 23: Take the modulus value of the real part and the imaginary part of the Numpy type data after removing the abnormal values, and obtain the data after the modulus value; Step 24: normalize the data after taking the modulus value to obtain the processed signal data.
3. The method for estimating strongly coherent multipath delay based on deep learning according to claim 2, characterized in that: In the step 3, the delay true value corresponding to the signal data in the data packet is subjected to sparse coding based on the Sinc kernel function to obtain the sparsely processed delay value; The specific process is: Step 31: Set the preset delay grid size and interval; Step 32: Find the delay grid closest to the true delay value corresponding to the signal data in the data packet, and thin out the delay value corresponding to the delay grid closest to the true delay value into a set of vectors consisting of 0 and 1; Step 3: Use Sinc to perform convolution calculation on the vector composed of 0 and 1 to obtain the convolution result p1; Step 34: perform weighted calculation on the actual delay value and the out-of-grid portion of the nearest preset grid value to obtain a weighted calculation result p2; Step 35: Add p1 and p2 to obtain the delay value after sparse processing.
4. The method for estimating strongly coherent multipath delay based on deep learning according to claim 3, characterized in that: In the step 32, the delay grid closest to the true delay value corresponding to the signal data in the data packet is found, and the delay value corresponding to the delay grid closest to the true delay value is sparsely converted into a set of vectors consisting of 0 and 1; the specific process is: in, is the preset delay grid value, δ( ) is the unit pulse function, D is the total number of signal paths, and i is the number of the i-th signal path.
5. The method for estimating strongly coherent multipath delay based on deep learning according to claim 4, characterized in that: In the step 33, Sinc is used to perform convolution calculation on the vector composed of 0 and 1 to obtain the convolution result p1; The calculation formula is as follows: in, The delay grid value closest to the actual delay value; is the convolution calculation, and sinc() is the sinc function.
6. The method for estimating strongly coherent multipath delay based on deep learning according to claim 5, characterized in that: In the steps 3 and 4, the actual delay value and the out-of-grid portion of the nearest preset grid value are weightedly calculated to obtain a weighted calculation result p2; The calculation formula is as follows: Among them, c is the set weighting coefficient, τ i It is the actual delay value of each path.
7. The method for estimating strongly coherent multipath delay based on deep learning according to claim 6, characterized in that: In the steps 3 and 5, p1 and p2 are added to obtain the delay value after sparse processing, and the formula is as follows: p=p1+p2 Where p is the delay value after sparse processing.
8. The method for estimating strongly coherent multipath delay based on deep learning according to claim 7, characterized in that: In the step 4, the data processed in the step 2 is used as the input of the delay regression network structure model, and the delay value after sparse processing obtained in the step 3 is used as the output of the delay regression network structure model, and the delay regression module network structure model is trained until convergence, so as to obtain a trained delay regression module network structure model; The specific process is: The network structure model of the time delay regression module includes the first fully connected layer FC, the first module, the second module, the third module, the fourth module, the fifth module, the sixth module, the seventh module, the eighth module, and the deep convolution layer; The working process of the network structure model of the time delay regression module is as follows: The data processed in step 1 is input into the first fully connected layer FC, and the first fully connected layer FC outputs feature A; Feature A is sequentially input into the first module, the second module, the third module, the fourth module, the fifth module, the sixth module, the seventh module, the eighth module, and the deep convolution layer, and the deep convolution layer outputs the feature; The output feature of the deep convolutional layer is the delay value after sparse processing obtained in step 3.
9. The method for estimating strongly coherent multipath delay based on deep learning according to claim 8, characterized in that: Each module in the first module, the second module, the third module, the fourth module, the fifth module, the sixth module, the seventh module, and the eighth module includes: a first convolutional layer, a first batch normalization BN, a first ReLuctant activation function layer, a second convolutional layer, a second batch normalization BN, a second ReLuctant activation function layer, a maximum pooling layer, a second fully connected layer FC, a third ReLuctant activation function layer, a third fully connected layer FC, a fourth ReLuctant activation function layer, and a multi-head attention mechanism layer.
10. The method for estimating strongly coherent multipath delay based on deep learning according to claim 9, characterized in that: The working process of each module in the first module, the second module, the third module, the fourth module, the fifth module, the sixth module, the seventh module and the eighth module is as follows: Feature A is sequentially input into the first convolutional layer, the first batch normalization BN, the first ReLuctant activation function layer, the second convolutional layer, the second batch normalization BN, the second ReLuctant activation function layer, and the second ReLuctant activation function layer outputs feature B; Feature B is sequentially input into the maximum pooling layer, the second fully connected layer FC, the third ReLuctant activation function layer, the third fully connected layer FC, the fourth ReLuctant activation function layer, and the fourth ReLuctant activation function layer outputs feature C; Feature B is input into the multi-head attention mechanism layer, and the multi-head attention mechanism layer outputs feature D; Multiply feature C and feature D element by element to obtain feature E; Feature E is added to feature A element by element to obtain feature F; Feature F is the output feature of each module.