A method for estimating time delay of off-grid direct wave based on deep learning algorithm

Abstract

The application relates to a method for estimating off-grid direct wave time delay based on a deep learning algorithm, and relates to the field of ocean resource development, in particular to a method for estimating off-grid direct wave time delay based on a deep learning algorithm. The application aims to solve the problem of poor precision of existing far-field direct wave time delay estimation. The process is as follows: active sonar transmits signal data, array element receives signal data, and signal data corresponds to time delay true value labels; processed signal data is obtained; a coarse estimation value of direct wave time delay is obtained; a difference value is calculated; a sparse processed time delay difference value is obtained; a trained time delay difference value regression module network structure model is obtained; array element received signal data is processed to obtain processed signal data; a feature vector is output based on the trained time delay difference value regression module network structure model; the feature vector is processed by using a peak searching function to obtain a time delay difference value; and the time delay difference value and the coarse estimation value of the direct wave time delay are added to obtain an estimated value of the received signal.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of marine resource development, and particularly relates to a method for off-grid direct wave time delay estimation based on a deep learning algorithm. BACKGROUND

[0002] Far-field direct wave time delay estimation is a research topic with wide application value in underwater positioning, and accurate direct sound time delay estimation is an important basis for target distance calculation, array calibration and positioning accuracy.

[0003] In the process of ultra-short baseline positioning, a sound beacon is installed on the underwater positioned target, the sound beacon emits a sound signal, and the ultra-short baseline system calculates the position and distance of the target after receiving the signal. Specifically, when the receiving device receives the signal, it records the time delay of the signal propagation to each receiving point, and calculates the position of the signal source. Existing methods include matching filter algorithm and model-driven and data-driven super-resolution algorithm.

[0004] The accuracy of the matching filter depends on the sampling rate of the data, and when the true value of the time delay is an integer multiple of the non-sampling point, the matching filter will have a large error. In the scene of high-precision direct wave time delay estimation, a more accurate algorithm is needed that is not constrained by the sampling point interval.

[0005] The model-driven method, such as the alternating direction multiplier method (ADMM), has some parameters that need to be selected manually, such as the step size parameter and the penalty function parameter. Different parameter selection may result in different convergence speed and stability of the algorithm, and some non-ideal conditions in reality, such as low signal-to-noise ratio, will cause the performance of the algorithm to deteriorate significantly.

[0006] The data-driven method is mainly based on deep learning time delay algorithm, but the traditional deep time delay estimation takes the true value of the TOA parameter as the learning target, and with the increase of network depth, the gradient will gradually disappear in the back propagation process. The overfitting problem caused by these factors will reduce the estimation accuracy of the test data. SUMMARY

[0007] The purpose of the present application is to solve the problem of poor accuracy of existing far-field direct wave time delay estimation, and to propose a method for off-grid direct wave time delay estimation based on a deep learning algorithm.

[0008] The specific process of a method for off-grid direct wave time delay estimation based on a deep learning algorithm is as follows:

[0009] Step 1: transmit signal data by active sonar, receive signal data by array element, and time delay true value label corresponding to signal data;

[0010] Step two, the signal data received by the array element and the active sonar emission signal data are sequentially subjected to data type conversion, abnormal value elimination, data real part and imaginary part module, data normalization and other processing, to obtain the processed signal data;

[0011] Step three, the signal data processed in step two is processed based on the matched filter algorithm to obtain a coarse estimation value of the direct wave time delay;

[0012] Step four, the difference between the time delay true value corresponding to the array element received signal data in step one and the coarse estimation value of the direct wave time delay obtained in step three is calculated;

[0013] Step five, the difference value in step four is sparsified based on the sparse coding mode of the Sinc kernel function to obtain a sparsified time delay difference value;

[0014] Step six, the signal data processed in step two is taken as the input of the time delay difference value regression network structure model, and the sparsified time delay difference value obtained in step five is taken as the output of the time delay difference value regression network structure model, the time delay difference value regression module network structure model is trained until convergence, and a trained time delay difference value regression module network structure model is obtained;

[0015] Step seven, the array element received signal data in step one is processed to obtain processed signal data;

[0016] The processed signal data is input into the trained time delay difference value regression module network structure model, and the trained time delay difference value regression module network structure model outputs a feature vector corresponding to each processed signal;

[0017] The feature vector output by the trained time delay regression module network structure model is processed by using a peak seeking function to obtain a time delay difference value;

[0018] The time delay difference value and the coarse estimation value of the direct wave time delay obtained in step three are added to obtain an estimated value of the received signal.

[0019] The beneficial effects of the present application are:

[0020] The present application introduces a deep convolutional network based on sparse coding on the basis of related algorithms. In the time delay estimation process, first, a coarse time delay value is calculated by using a correlation processor, and then a deep convolutional network based on sparse coding is used to train and predict the difference between the coarse time delay value and the true time delay value;

[0021] The present application can break through the constraint of data sampling rate, accurately estimate the time delay of non-integer multiple sampling points, and has more accurate estimation effect at a lower signal-to-noise ratio;

[0022] The residual structure of the application can make the network better capture and process deeper features by introducing a skip connection. At the same time, by adding an input to the function output, this operation can be explicitly split into two linear superpositions, making learning easier and alleviating the training problem of deep networks;

[0023] In each algorithm module of the application, two convolutional layers, one pooling layer, two fully connected layers and one residual combination are combined. The processing layer composed of convolutional layer, pooling layer and fully connected layer learns the weight of each feature, and the normalization layer is added to the residual block to enhance the stability of the features in the network. The features are compressed by the pooling layer to obtain the global receptive field. Then these features are input to two fully connected layers. Finally, the residual structure is used to avoid the vanishing features during transmission between deep layers. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The overall algorithm flowchart is shown in the figure;

[0025] Figure 2 The data preprocessing module flowchart is shown in the figure;

[0026] Figure 3 The matched filter calculation flowchart is shown in the figure;

[0027] Figure 4 The time delay regression module network structure diagram is shown in the figure;

[0028] Figure 5 The training set loss change diagram of 10dB is shown in the figure;

[0029] Figure 6 The matched filter algorithm and the application under different signal-to-noise ratio delay estimation mean square error curve comparison diagram is shown in the figure;

[0030] Figure 7 The delay estimation accuracy (confidence interval is 2ns) of the application under different signal-to-noise ratio is shown in the figure;

[0031] Figure 8 The schematic diagram of the simulation signal data of the application obtained by mixing two signals is shown in the figure;

[0032] Figure 9 The time delay grid diagram is shown in the figure. DETAILED DESCRIPTION

[0033] Specific implementation one: the specific process of the direct wave time delay estimation method based on deep learning algorithm in this embodiment is as follows:

[0034] Step one, active sonar signal data, array element received signal data and signal data corresponding time delay true value label;

[0035] Step two, the signal data received by the array element and the active sonar transmitting signal data are sequentially subjected to data type conversion, abnormal value elimination, data real part and imaginary part module, data normalization and other processing, to obtain the processed signal data;

[0036] Step three, the signal data processed in step two is processed based on a matched filter algorithm to obtain a coarse estimation value of the direct wave time delay;

[0037] Taking two-path signal as an example, each simulation signal data is obtained by mixing two signals, and the corresponding two labels are, for example, Figure 8 ;

[0038] Suppose signal 1 (red), the real time delay difference value is 1x10 -5 s, and the corresponding time delay grid is 1.

[0039] Suppose signal 2 (blue), the real time delay difference value is 4x10 -5 s, and the corresponding time delay grid is 4.

[0040] Two signals are superimposed to obtain a simulation signal, the label is (1x10 -5 , 4x10 -5 ), and the corresponding grid value is (1, 4); the time delay grid is shown in Figure 9 .

[0041] The active sonar transmitting signal data is used to construct a filter in the matched filter algorithm.

[0042] The present application is based on the combination of time delay coarse estimation module and time delay regression module: for far-field time delay estimation, if a deep learning algorithm using sparse coding is used to estimate the time delay, grid points covering a large time domain range need to be set, so a single sparse coding method is obviously not suitable for far-field multi-path time delay estimation; therefore, a coarse-precision estimation process is proposed.

[0043] The time delay estimation based on matched filtering maximizes the peak value of the signal by designing a filter matched with the expected signal and using the correlation between the signal and the filter, so as to realize accurate estimation of the signal delay. By observing the peak position of the matched filter output, the time corresponding to the peak position of the signal delay can be determined as the estimated delay of the signal.

[0044] Step four, calculate the difference between the time delay true value corresponding to the array element signal data in step one and the coarse estimation value of the direct wave time delay obtained in step three.

[0045] Step five, the difference value in step four is subjected to sparse processing based on the sparse coding mode of Sinc kernel function, to obtain the sparse processed time delay difference value.

[0046] Step six, the signal data processed in step two is taken as the input of the time delay difference value regression network structure model, the time delay difference value processed in step five is taken as the output of the time delay difference value regression network structure model, the time delay difference value regression module network structure model is trained until convergence, and the trained time delay difference value regression module network structure model is obtained;

[0047] Step seven, the array element received signal data (only data, without corresponding labels) in step one is processed to obtain processed signal data;

[0048] The processed signal data is input into the trained time delay difference value regression module network structure model, and the trained time delay difference value regression module network structure model outputs the feature vector corresponding to each processed signal;

[0049] The feature vector output by the trained time delay regression module network structure model is processed by using a peak seeking function to obtain a time delay difference value;

[0050] 200 data, each data has 1 time delay label, and each data has 1 peak value;

[0051] The time delay difference value and the rough estimation value of the direct wave time delay obtained in step three are added to obtain a high-precision estimation value of the received signal.

[0052] Through the application, the constraint of the time domain sampling frequency can be broken, and high-precision time delay estimation can be performed on a non-time domain sampling grid.

[0053] Specific implementation method two: different from the specific implementation method one, the signal data received by the array element and the active sonar transmitted signal data in step two are sequentially subjected to data type conversion, abnormal value elimination, data real part and imaginary part module, data normalization and other processing to obtain processed signal data.

[0054] The specific process is as follows:

[0055] Step two one, the signal data received by the array element is subjected to data type conversion to obtain Numpy type data;

[0056] Step two two, the obtained Numpy type data is subjected to abnormal value elimination (abnormal value refers to missing data);

[0057] Step two three, the real part and the imaginary part of the Numpy type data after the abnormal value elimination are subjected to module value to obtain data after the module value;

[0058] Step two four, the data after the module value is subjected to normalization processing to obtain processed signal data.

[0059] The other steps and parameters are the same as those in the specific implementation method one.

[0060] Specific implementation three: the difference value of step four is sparsely processed in step five based on the Sinc kernel function to obtain a sparsely processed time delay difference value;

[0061] The specific process is:

[0062] The Sinc function is used to process the function value corresponding to the time delay grid to obtain the code with more sidelobe information, and the steps are as shown in Figure 2 ;

[0063] Step five one, set the preset time delay grid size and interval;

[0064] The time delay grid abscissa: the preset time delay grid range is 0-10 (including 0 and 10), and the corresponding real time domain range is 0-10 -5 s (including 0 and 10), the time delay grid in the range of 0-10 is divided into 1000 small grids (that is, the grid resolution (adjacent grid interval) is set to 0.01), and each small grid corresponds to a real time domain of 10 -8 s (grid value);

[0065] 1000 small grids mean that a time delay vector is composed of 1000 0 or 1 arrays;

[0066] The time delay grid ordinate: the sparsely processed time delay value;

[0067] Step five two, find the nearest preset time delay grid to the difference value obtained in step four, and sparsely process the time delay difference value (such as 0.1) corresponding to the nearest preset time delay grid to the difference value (such as 0.11) obtained in step four into a vector composed of 0 and 1;

[0068] Step five three, use Sinc to perform convolution calculation on the vector composed of 0 and 1 to obtain a convolution result p1;

[0069] Step five four, perform weighted calculation on the off-grid part (such as 0.01) of the time delay difference value (such as 0.1) corresponding to the nearest preset time delay grid to the difference value (such as 0.11) obtained in step four and the difference value (such as 0.11) obtained in step four to obtain a weighted calculation result p2;

[0070] Step five five, add p1 and p2 to obtain a sparsely processed time delay difference value.

[0071] The sparsely processed time delay difference value is used as a training label of a deep learning input signal;

[0072] The true value of the TOA parameter is represented using sparse coding, and the network is optimized by minimizing the error between the output of the training network and the sparse representation of the training TOA.

[0073] The sparse coding module is based on a Sinc kernel function: after processing the real time delay value corresponding to the time delay grid using a Sinc function, the sparse coding with more sidelobe information can be obtained, so that the present application has more superior accuracy and stability.

[0074] The other steps and parameters are the same as those in embodiment one or two.

[0075] Embodiment four: different from one of embodiments one to three, in step five two, the preset time delay grid closest to the sparse processed time delay difference value obtained in step three is found, and the time delay difference value (such as 0.1) corresponding to the preset time delay grid closest to the sparse processed time delay difference value (such as 0.11) obtained in step three is sparse to a vector composed of 0 and 1.

[0076] The specific process is as follows:

[0077]

[0078] wherein, is the preset time delay grid value (such as 0.1) closest to the sparse processed time delay difference value (such as 0.11) obtained in step three, is the preset time delay grid value, δ( ) is the unit impulse function, D is the total number of signal paths, and i is the i-th signal path.

[0079] The other steps and parameters are the same as those in one of embodiments one to three.

[0080] Embodiment five: different from one of embodiments one to four, in step five three, a vector composed of 0 and 1 is calculated by convolution using Sinc, and a convolution result p1 is obtained.

[0081] The calculation formula is as follows:

[0082]

[0083] wherein, is the convolution calculation, and sinc( ) is the Sinc function.

[0084] The other steps and parameters are the same as those in one of embodiments one to four.

[0085] Specific implementation six: different from one of the specific implementations one to five, the step five four is to calculate the weighted calculation result p2 by weighting the off-grid part (such as 0.11-0.1=0.01) of the delay difference value (such as 0.1) corresponding to the preset delay grid closest to the sparse processed delay difference value (such as 0.11) obtained in step three in step five three.

[0086]

[0087] Wherein, c is the set of weighted coefficients, τ i is the real delay value of each path.

[0088] The other steps and parameters are the same as one of the specific implementations one to five.

[0089] Specific implementation seven: different from one of the specific implementations one to six, the step five five is to add p1 and p2 to obtain the sparse processed delay value, the formula is as follows:

[0090] p=p1+p2

[0091] Wherein, p is the sparse processed delay value.

[0092] The other steps and parameters are the same as one of the specific implementations one to six.

[0093] Specific implementation eight: different from one of the specific implementations one to seven, the step six is to take the signal data processed in step two as the input of the delay difference value regression network structure model, take the sparse processed delay difference value obtained in step five as the output of the delay difference value regression network structure model, train the delay difference value regression module network structure model until convergence, and obtain the trained delay difference value regression module network structure model;

[0094] The specific process is as follows:

[0095] The delay difference value regression module network structure model includes a first full connection layer FC, a first module, a second module, a third module, and a deep convolution layer.

[0096] The working process of the delay difference value regression module network structure model is as follows:

[0097] The signal data processed in step two is input into the first full connection layer FC, and the first full connection layer FC outputs the feature A.

[0098] The feature A is input into the first module, the second module, and the third module in sequence, and the third module outputs the feature A'.

[0099] The third module outputs a feature A' and performs element-wise sum with the first fully connected layer FC output feature A to obtain a feature A'';

[0100] The feature A'' is input into a deep convolutional layer, and the deep convolutional layer outputs a feature.

[0101] The deep convolutional layer output feature is the time delay difference value obtained after sparse processing in step five.

[0102] The other steps and parameters are the same as one of the first to seventh embodiments.

[0103] The ninth embodiment is different from one of the first to eighth embodiments in that each of the first module, the second module and the third module comprises: 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 max-pooling layer, a second fully connected layer FC, a third ReLuctant activation function layer, a third fully connected layer FC and a fourth ReLuctant activation function layer.

[0104] The other steps and parameters are the same as one of the first to eighth embodiments.

[0105] The tenth embodiment is different from one of the first to ninth embodiments in that the working process of each of the first module, the second module and the third module is as follows:

[0106] The 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 and the second ReLuctant activation function layer, and the second ReLuctant activation function layer outputs a feature B.

[0107] The feature B is sequentially input into the max-pooling layer, the second fully connected layer FC, the third ReLuctant activation function layer, the third fully connected layer FC and the fourth ReLuctant activation function layer, and the fourth ReLuctant activation function layer outputs a feature C.

[0108] The feature C is the output feature of each module.

[0109] The other steps and parameters are the same as one of the first to ninth embodiments.

[0110] 1) Convolution operation: by sliding a small filter over the input data, the dot product of the filter and the local region of the input data is calculated to obtain 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.

[0111] 2) Activation function, used to introduce nonlinearity, so that the model can learn more complex features. The role of these functions is to enhance the expression ability of the model, so that it can handle a wider range of input data.

[0112] 3) Pooling layer, used to reduce the dimension of the feature map, reduce the amount of calculation, while retaining important feature information. Common pooling methods include max pooling and average pooling. The role of the pooling layer is to further extract and concentrate features, so that the model has better robustness to small changes in input data.

[0113] 4) Fully connected layer, which integrates the features extracted by the previous convolutional and pooling layers, performs linear transformation through weights and biases, and then performs nonlinear transformation through activation functions to finally get the classification result. The role of the fully connected layer is to map the learned features to the sample label space, completing the classification or regression task.

[0114] 5) Normalization layer, which standardizes the output of each layer to speed up training and improve model stability. Common ones include batch normalization and layer normalization.

[0115] In the time delay difference value regression module of the application, a residual structure is introduced: in order to reduce the risk of gradient disappearance and overfitting in the training process of deep neural networks, a residual structure is introduced. The residual structure can better capture and process deeper features through the introduction of skip connections. With the increase of network depth, the gradient will gradually disappear in the backpropagation process, making it difficult to train the network. The residual structure directly connects the input to the output through a skip connection, so that the gradient can flow back to the input layer directly, thereby alleviating the problem of gradient disappearance. At the same time, by adding an input to the function output, the relationship between the input and the output can still be described, which can explicitly split into two linear superpositions, making learning easier and alleviating the training problem of deep networks, providing the possibility for training thousands of layers of networks.

[0116] In the time delay difference value regression module of the application, a deep convolutional network with a residual structure is encapsulated as an algorithm module: in each algorithm module, there are two convolutional layers, a pooling layer, two fully connected layers, and a residual structure. The specific structure is as follows:

[0117] Structure 1) Two convolutional layers are used to learn the weights of each feature channel and extract TOA features.

[0118] Structure 2) A pooling layer is used to compress features to obtain a global receptive field.

[0119] Structure 3) Two fully connected layers are used to calculate the nonlinear relationship of compressed feature channels, where the activation function is selected as the Relu function.

[0120] Structure 4) uses residual blocks to avoid losing a part of TOA features during transmission between deep layers.

[0121] The above structure is encapsulated as an algorithm module, and when preset parameters are changed, the number of algorithm modules can be directly changed to quickly increase the number of layers of the neural network, and the residual structure can effectively avoid the occurrence of overfitting and improve the performance of network feature extraction.

[0122] Embodiments:

[0123] The off-grid direct wave time delay estimation method based on the deep learning algorithm of the embodiment of the application is implemented as shown in Figure 1 , and specifically as follows:

[0124] Step 1), in the scene of far-field direct wave time delay estimation, the correlation algorithm, the correlation interpolation method and the proposed deep learning time delay estimation algorithm based on sparse coding are simulated, and the simulation signal parameters are as follows: signal length 10 ms, frequency range 20 kHz-30 kHz, sampling frequency 100 kHz, array element number 1, signal time delay value 100*10 -5 s to 150*10 -5 s.

[0125] Step 2), set the required parameters in the algorithm module, collect 100,000 samples, randomly divide them into 90% for training and 10% for verification, the background noise is white Gaussian noise with additive noise corresponding to the sampling signal, and the training is based on Pytorch and Adam optimizer. The real TOA values of the training set and the test set are randomly generated in the specified interval. For the algorithm containing sparse coding, the length of the pre-defined over-complete dictionary is 1000. The root mean square error (RMSE) of all TOA estimation values is calculated to evaluate the estimation accuracy.

[0126] Step 3), the data preprocessing module is used for the following processing, and the implementation steps are as shown in Figure 2 , and specifically as follows:

[0127] Step 3-1), type conversion is performed on the array element received data and the active sonar transmission signal data to obtain Numpy type data;

[0128] Step 3-2), the data is encapsulated into a three-dimensional array with a dimension of 200*2*1200 for each batch of training data, wherein the received signal is the first row data of the second dimension, and the transmission signal is the second row data of the second dimension;

[0129] Step 3-3), the converted type data is subjected to outlier rejection to exclude the interference of outliers on signal processing;

[0130] Step 3-4), taking the modulus of the real part and the imaginary part of the processed data and performing normalization processing to obtain standard data.

[0131] Step 4), calculating the coarse estimate value of the time delay by using the time delay coarse estimation module, the implementation steps are as shown in Figure 3 , and the specific implementation steps are as follows:

[0132] Step 4-1), constructing a matched filter of the expected signal by using the transmitted signal data in the input array.

[0133] Step 4-2), correlating the data of the received signal in the input array with the constructed matched filter.

[0134] Step 4-3), finding the peak position in the correlated data and subtracting the value of half of the preset time delay estimation grid length as the output value of the time delay coarse estimation module.

[0135] Step 4-4), deleting the transmitted signal data in the last input array.

[0136] Step 5), calculating the label of the time delay estimation by using the sparse coding module, and the specific steps are as follows:

[0137] Step 5-1), subtracting the coarse estimate value output by the time delay estimation coarse estimation module from the true time delay value to obtain a residual value.

[0138] Step 5-2), processing the function value corresponding to the time delay grid by using the Sinc function to obtain a code with more sidelobe information, and the off-grid value is calculated as an additional value at the nearest grid position. Similarly, the off-grid time delay can be calculated according to the additional value.

[0139] Step 6), inputting the sparse coded label and the signal data output by the time delay coarse estimation module into the network structure of the time delay regression module for training, and the steps are as shown in Figure 4 , and the specific implementation steps are as follows:

[0140] Step 6-1), building a deep learning network structure according to the preset parameters, and the network structure is composed of a plurality of algorithm modules, each of which is composed of two convolution layers, one pooling layer, two fully connected layers, and a residual structure. The specific structure is as follows:

[0141] Structure 1), learning the weight of each feature channel by using two convolution layers to extract the TOA feature.

[0142] Structure 2), compressing the feature by using one pooling layer to obtain a global receptive field.

[0143] Structure 3), calculating the nonlinear relationship of the compressed feature channel by using two fully connected layers, and the activation function is selected as the Relu function.

[0144] Structure 4), using residual blocks to avoid a portion of TOA features disappearing during transmission between deep layers.

[0145] When setting the preset parameters, the number of modules is 3, there are 6 convolutional layers, 3 pooling layers, 6 fully connected layers and 3 residual structures.

[0146] In the present application, the convolution kernel size, step size and other hyperparameters of each convolutional layer and fully connected layer are as shown in Figure 4 .

[0147] Step 6-2), randomly set the weights and bias values of the input layer, and establish the non-linear relationship between the layers in the forward propagation process, wherein in the process of forward propagation, the input data is processed layer by layer through the input layer and the hidden layer and transmitted to the output layer.

[0148] Step 6-3), calculate the sum of squares of the error values of the output layer and the set label as a constraint function, and use the target function to calculate the partial derivative of each neuron weight in the back propagation, and gradually correct each weight value using the stochastic gradient descent algorithm.

[0149] Step 6-4), repeat steps 6-2) and 6-3) until the training number reaches the preset value.

[0150] Step 6-5), Figure 5 is the loss change graph of the neural network training set when the signal-to-noise ratio is set to 10dB.

[0151] Step 7), repeat steps 1) to step 6) under multiple signal-to-noise ratios when the signal-to-noise ratio is in the range of -5 to 10dB with a step size of 3dB, and obtain the mean square error curve of the delay estimation of the present application under different signal-to-noise ratios as shown in Figure 6 , and calculate the delay estimation accuracy under different signal-to-noise ratios when the confidence interval is 0.25 adjacent sampling point intervals as shown in Figure 7 .

[0152] Step 8), as shown in Figure 6 , it can be seen that when the signal-to-noise ratio is -5 to 10dB with a step size of 3dB, the RMSE value of the algorithm of the present application is significantly reduced compared with the matching filter algorithm. That is, in the far-field direct wave delay estimation scenario, the delay estimation accuracy of the algorithm proposed in this paper is significantly improved compared with the related delay estimation algorithm. When SNR<10dB, the RMSE value of the algorithm decreases by more than 66%

[0153] Step 9), as shown in Figure 7 , it can be seen that when the signal-to-noise ratio is -5 to 10dB, the delay estimation accuracy of the algorithm of the present application is high, all above 75%. The delay estimation accuracy of the matching filter algorithm is all below 45%.

[0154] The present application can have other various embodiments, and those skilled in the art can make various corresponding changes and modifications according to the present application without departing from the spirit and essence of the present application, and these corresponding changes and modifications shall all belong to the protection scope of the claims of the present application.

Claims

1. A method for off-grid direct wave time delay estimation based on deep learning algorithm, characterized in that: The method specifically comprises the following steps: Step one, active sonar signal data transmission, array element received signal data and signal data corresponding to the time delay true value label; Step two, sequentially performing data type conversion, abnormal value elimination, data real part and imaginary part module, and data normalization processing on the signal data received by the array element and the active sonar signal data transmission, to obtain processed signal data; Step three, processing the signal data processed in step two based on a matched filter algorithm to obtain a rough estimation value of the direct wave time delay; Step four, calculating the difference between the time delay true value corresponding to the array element received signal data in step one and the rough estimation value of the direct wave time delay obtained in step three; Step five, performing sparse processing on the difference value in step four based on a sparse coding mode of a Sinc kernel function to obtain a sparse processed time delay difference value; The specific process comprises the following steps: Step five one, setting a preset time delay grid size and interval; Step five two, finding the nearest preset time delay grid to the difference value obtained in step four, and sparsely processing the time delay difference value corresponding to the nearest preset time delay grid to the difference value obtained in step four into a vector composed of 0 and 1; Step five three, performing convolution calculation on the vector composed of 0 and 1 by using Sinc to obtain a convolution result p1; Step five four, performing weighted calculation on the difference value obtained in step four and the off-grid part of the time delay difference value corresponding to the nearest preset time delay grid to the difference value obtained in step four to obtain a weighted calculation result p2; Step five five, adding p1 and p2 to obtain the sparse processed time delay difference value; Step six, taking the signal data processed in step two as the input of a time delay difference value regression network structure model, taking the sparse processed time delay difference value obtained in step five as the output of the time delay difference value regression network structure model, training the time delay difference value regression module network structure model until convergence, and obtaining a trained time delay difference value regression module network structure model; Step seven, processing the array element received signal data in step one to obtain processed signal data; Inputting the processed signal data into the trained time delay difference value regression module network structure model, and outputting a feature vector corresponding to each processed signal by the trained time delay difference value regression module network structure model; Processing the feature vector output by the trained time delay regression module network structure model by using a peak searching function to obtain a time delay difference value; Adding the time delay difference value and the rough estimation value of the direct wave time delay obtained in step three to obtain an estimated value of the received signal.

2. The method of claim 1, wherein the method is based on a deep learning algorithm. In step two, the signal data received by the array element and the active sonar signal data transmission are sequentially subjected to data type conversion, abnormal value elimination, data real part and imaginary part module, and data normalization processing to obtain processed signal data; The specific process comprises the following steps: Step two one, performing data type conversion on the signal data received by the array element to obtain Numpy type data; Step two two, performing abnormal value elimination on the obtained Numpy type data; Step two three, taking the module value of the real part and the imaginary part of the Numpy type data after the abnormal value elimination to obtain data after the module value; Step two four, performing normalization processing on the data after the module value to obtain processed signal data.

3. The method of claim 2, wherein the method is based on a deep learning algorithm. The step five two finds the preset delay grid closest to the sparse processed delay difference value obtained in step three, and sparsifies the delay difference value corresponding to the preset delay grid closest to the sparse processed delay difference value obtained in step three into a vector composed of 0 and 1; The specific process is as follows: wherein, is the preset delay grid value closest to the sparse-processed delay difference value obtained in step three, is the preset delay grid value, δ( ) is the unit impulse function, D is the total number of signal paths, and i is the i-th signal path number.

4. The method of claim 3, wherein the method is based on a deep learning algorithm. The step five three performs convolution calculation on the vector composed of 0 and 1 by using Sinc to obtain a convolution result p1; The calculation formula is as follows: wherein is a convolution calculation, and sinc() is a sine function.

5. The method of claim 4, wherein the method is based on a deep learning algorithm. The step five four performs weighted calculation on the off-grid part of the delay difference value corresponding to the preset delay grid closest to the sparse processed delay difference value obtained in step three, and the delay difference value obtained in step three to obtain a weighted calculation result p2; The calculation formula is as follows: where c is a set weighting factor, τ i is the per-path true time delay value.

6. The method of claim 5, wherein the method is based on a deep learning algorithm. The step five five adds p1 and p2 to obtain a sparse processed delay value; The formula is as follows: p = p1 + p2 Wherein, p is the sparse processed delay value.

7. The method of claim 6, wherein the method is based on a deep learning algorithm. The step six takes the signal data processed in step two as the input of the delay difference value regression network structure model, takes the sparse processed delay difference value obtained in step five as the output of the delay difference value regression network structure model, trains the delay difference value regression module network structure model until convergence, and obtains the trained delay difference value regression module network structure model; The specific process is as follows: The delay difference value regression module network structure model comprises a first full connection layer FC, a first module, a second module, a third module and a deep convolution layer. The working process of the delay difference value regression module network structure model is as follows: The signal data processed in step two is input into the first full connection layer FC, and the first full connection layer FC outputs a feature A; The feature A is sequentially input into the first module, the second module and the third module, and the third module outputs a feature A'; The feature A' output by the third module and the feature A output by the first full connection layer FC are element-wise summed to obtain a feature A"; The feature A" is input into the deep convolution layer, and the deep convolution layer outputs a feature; The feature output by the deep convolution layer is the sparse processed delay difference value obtained in step five.

8. The method of claim 7, wherein the method is based on a deep learning algorithm. Each of the first module, the second module and the third module comprises a first convolution layer, a first batch normalization BN, a first ReLuctant activation function layer, a second convolution layer, a second batch normalization BN, a second ReLuctant activation function layer, a maximum pooling layer, a second full connection layer FC, a third ReLuctant activation function layer, a third full connection layer FC and a fourth ReLuctant activation function layer.

9. The method of claim 8, wherein the method is based on a deep learning algorithm. The working process of each of the first module, the second module and the third module is as follows: The feature A is sequentially input into the first convolution layer, the first batch normalization BN, the first ReLuctant activation function layer, the second convolution layer, the second batch normalization BN and the second ReLuctant activation function layer, and the second ReLuctant activation function layer outputs a feature B; The feature B is sequentially input into the maximum pooling layer, the second full connection layer FC, the third ReLuctant activation function layer, the third full connection layer FC and the fourth ReLuctant activation function layer, and the fourth ReLuctant activation function layer outputs a feature C; The feature C is the output feature of each module.

Images