High-precision delay estimation method and system for broadband digital transmission signals based on CNN

Abstract

The present application discloses a high-precision delay estimation method and system for broadband digital transmission signals based on CNN. The method comprises the following steps: obtaining multiple groups of discrete target signals in complex time domain form received by multiple antennas; performing channel simulation on each group of target signals to obtain simulated target signals with added phase error and noise; processing a reference signal and a simulated target signal according to a phase estimation model to obtain a predicted phase difference of the simulated target signal relative to the reference signal, wherein the reference signal is any one of the multiple groups of simulated target signals, and the phase estimation model is a convolutional neural network model; determining the delay difference and initial phase difference of the simulated target signal relative to the reference signal according to the predicted phase difference. The scheme uses the CNN model to estimate the delay difference and initial phase difference of the signal, which can significantly improve the accuracy and speed of delay estimation between broadband digital transmission signals under low signal-to-noise ratio, reduce the amount of calculation, and enhance real-time performance.

Description

Technical Field

[0001] The present application relates to the field of digital communication technology, and in particular to a method and system for high-precision delay estimation of broadband digital transmission signals based on CNN. Background Art

[0002] In deep space antenna array systems, due to the long distance, the signal power reaching the ground in deep space communication is very weak, and the output signal-to-noise ratio of each unit of the antenna array is usually very low. The existing delay estimation methods based on cross-correlation (such as time domain cross-correlation, frequency domain cross-correlation, generalized cross-correlation, quadratic cross-correlation, etc.) will produce large errors or even fail completely, greatly reducing the signal synthesis performance of the antenna array. For example, the principle of the method based on frequency domain cross-correlation is to convert the time domain signals of two groups of antennas into the frequency domain, perform frequency domain cross-correlation, and then take the inverse tangent to obtain the warped phase difference of the two groups of antennas, unwrap the warped phase difference to obtain the true phase difference, and finally fit the "true phase difference-frequency" curve to obtain the slope and intercept to further obtain the delay value. However, under extremely low signal-to-noise ratio conditions, due to the influence of background noise, the phase information of the signal received by the antenna is easily interfered, and the signal will be submerged in the noise. The use of commonly used cross-correlation algorithms will cause the accuracy of delay estimation to drop significantly, resulting in a large delay error in the subsequent delay. In addition, the cross-correlation-based method is extremely computationally intensive. Most of the time spent on the entire antenna array signal synthesis system comes from cross-correlation calculations. As the size of the antenna array increases, the complexity of signal processing increases exponentially, which directly leads to a sharp increase in the amount of algorithm calculations. High computing requirements not only consume a large amount of processor resources, but also a large number of computing tasks slow down data processing speed and reduce the system's real-time response capabilities. Summary of the invention

[0003] To this end, this application discloses the following technical solutions:

[0004] The first aspect of the present application provides a high-precision delay estimation method for broadband digital transmission signals based on CNN, comprising:

[0005] Acquire multiple groups of target signals, each group of target signals is received by one antenna, and the target signals are discrete signals in complex time domain form;

[0006] Performing channel simulation on each group of the target signals to obtain simulated target signals with added phase error and noise;

[0007] Processing a reference signal and each of the simulated target signals except the reference signal according to a phase estimation model to obtain a predicted phase difference of each of the simulated target signals except the reference signal relative to the reference signal, wherein the reference signal is any one of a plurality of groups of simulated target signals, and the phase estimation model is a convolutional neural network model;

[0008] According to the predicted phase difference corresponding to each of the simulated target signals except the reference signal, the time delay difference and initial phase difference of each of the simulated target signals except the reference signal relative to the reference signal are determined.

[0009] Optionally, also include:

[0010] Obtaining a time delay difference and an initial phase difference of the reference signal relative to the reference signal;

[0011] The performance of the phase estimation model is detected according to the time delay difference and the initial phase difference of the reference signal relative to the reference signal.

[0012] Optionally, the detecting the performance of the phase estimation model according to the delay difference and the initial phase difference of the reference signal relative to the reference signal includes:

[0013] When the delay difference and the initial phase difference of the reference signal relative to the reference signal are both equal to 0, determining that the performance of the phase estimation model is normal;

[0014] When at least one of the time delay difference and the initial phase difference of the reference signal relative to the reference signal is not equal to 0, it is determined that the performance of the phase estimation model is abnormal.

[0015] Optionally, the method for obtaining the phase estimation model includes:

[0016] Processing samples with the model to be trained to obtain predicted phase differences corresponding to the samples, wherein the number of the samples is multiple, and each of the samples includes two groups of input signals and a group of labels;

[0017] Wherein, each group of the input signals of the sample is a discrete signal in complex time domain form, and the label of the sample includes label data corresponding to the sampling points of the input signal one by one, and the label data is the real phase difference between the two groups of input signals of the sample at the corresponding sampling points;

[0018] The model to be trained is trained according to the predicted phase difference and the true phase difference of the sample until a training condition is met to obtain the phase estimation model.

[0019] Optionally, before processing the sample with the model to be trained, the method further includes:

[0020] Preprocessing the sample to obtain a preprocessed sample;

[0021] The processing of samples using the model to be trained includes:

[0022] Processing the preprocessed samples with the model to be trained;

[0023] The preprocessing includes at least one of data normalization, data division and format conversion.

[0024] The second aspect of the present application provides a high-precision delay estimation system for broadband digital transmission signals based on CNN, comprising:

[0025] A signal source module, used to obtain multiple groups of target signals, each group of target signals is received by one antenna, and the target signals are discrete signals in complex time domain form;

[0026] A simulation module, used for performing channel simulation on each group of target signals to obtain simulated target signals with added phase error and noise;

[0027] an estimation module, configured to process a reference signal and each of the simulated target signals except the reference signal according to a phase estimation model to obtain a predicted phase difference of each of the simulated target signals except the reference signal relative to the reference signal, wherein the reference signal is any one of a plurality of groups of the simulated target signals, and the phase estimation model is a convolutional neural network model;

[0028] A determination module is used to determine the time delay difference and initial phase difference of each of the simulated target signals except the reference signal relative to the reference signal based on the predicted phase difference corresponding to each of the simulated target signals except the reference signal.

[0029] Optionally, the determining module is further used for:

[0030] Obtaining a time delay difference and an initial phase difference of the reference signal relative to the reference signal;

[0031] The performance of the phase estimation model is detected according to the time delay difference and the initial phase difference of the reference signal relative to the reference signal.

[0032] Optionally, when the determination module detects the performance of the phase estimation model according to the delay difference and the initial phase difference of the reference signal relative to the reference signal, it is used to:

[0033] When the delay difference and the initial phase difference of the reference signal relative to the reference signal are both equal to 0, determining that the performance of the phase estimation model is normal;

[0034] When at least one of the time delay difference and the initial phase difference of the reference signal relative to the reference signal is not equal to 0, it is determined that the performance of the phase estimation model is abnormal.

[0035] Optionally, a training module is also included to:

[0036] Processing samples with the model to be trained to obtain predicted phase differences corresponding to the samples, wherein the number of the samples is multiple, and each of the samples includes two groups of input signals and a group of labels;

[0037] Wherein, each group of the input signals of the sample is a discrete signal in complex time domain form, and the label of the sample includes label data corresponding to the sampling points of the input signal one by one, and the label data is the real phase difference between the two groups of input signals of the sample at the corresponding sampling points;

[0038] The model to be trained is trained according to the predicted phase difference and the true phase difference of the sample until a training condition is met, so as to obtain the phase estimation model.

[0039] Optionally, the training module is also used to:

[0040] Preprocessing the sample to obtain a preprocessed sample;

[0041] The processing of samples using the model to be trained includes:

[0042] Processing the preprocessed samples with the model to be trained;

[0043] The preprocessing includes at least one of data normalization, data division and format conversion.

[0044] The beneficial effects of this solution are:

[0045] Using the CNN model to estimate the delay difference and initial phase difference of the signal can significantly improve the accuracy of delay estimation between broadband digital transmission signals under low signal-to-noise ratio, and can significantly improve the speed of delay estimation between broadband digital transmission signals under low signal-to-noise ratio, reduce the amount of calculation, and enhance real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0047] Figure 1 is a structural schematic diagram of a phase estimation model provided in an embodiment of the present application;

[0048] Figure 2 It is a schematic diagram of a curve showing a change in a model loss value during a training process and a model loss value during a verification process provided by an embodiment of the present application;

[0049] Figure 3It is a schematic diagram of comparison between the predicted phase difference and the actual phase difference of a phase estimation model obtained through training on a validation set provided by an embodiment of the present application;

[0050] Figure 4 It is a flow chart of a high-precision delay estimation method for broadband digital transmission signals based on CNN provided in an embodiment of the present application;

[0051] Figure 5 It is a structural diagram of a high-precision delay estimation system for broadband digital transmission signals based on CNN provided in an embodiment of the present application. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0053] In order to apply the method of this embodiment, the following model data set may be first constructed so as to use the model data set to train and verify the phase estimation model used by the method of this embodiment.

[0054] The model data set may include multiple samples, each of which may include two sets of input signals X1 and X2, and a set of labels Y.

[0055] For each sample, the two groups of input signals in one sample may be two groups of antenna broadband digital transmission signals. For example, the input signal X1 may be the antenna broadband digital transmission signal received by antenna 1 , and the input signal X2 may be the antenna broadband digital transmission signal received by antenna 2 .

[0056] Each group of antenna broadband digital transmission signals can be represented in the form of complex time domain signals. When storing samples, a separate storage strategy can be adopted for the real and imaginary parts of each group of antenna broadband digital transmission signals. Each group of input signals of a sample can contain 2000 sampling points, that is, each group of antenna broadband digital transmission signals of the sample can be represented by discrete data obtained by sampling at 2000 sampling points.

[0057] In some embodiments, a set of input signals may also include other numbers of sampling points, such as 3000, 1000, etc., and is not limited to 2000 sampling points.

[0058] As an example, the two groups of input signals contained in a sample can be represented by the following Table 1. It should be noted that the header (first row and first column) in Table 1 is only for explanation. The input signal X of each sample in the actual model data set does not include the header, and the "data" in Table 1 is only used as an example. The actual data shall be based on the collected signal.

[0059] Table 1

[0060] Sampling point number Real part of X1 Imaginary part of X1 Real part of X2 Imaginary part of X2 1 data data data data 2 data data data data … … … … … 2000 data data data data

[0061] The label Y contains the same number of label data as the sampling points of the input signal. Combined with the above example, a set of label Y can contain 2000 label data. The label data and the sampling points of the input signal correspond one to one. Each label data can represent the actual phase difference between the two groups of input signals at the corresponding sampling point, in radians.

[0062] As an example, a set of labels Y contained in the above sample can be represented by Table 2, where data 1 can represent the true phase difference between the two signals X1 and X2 in Table 1 at sampling point 1, data 2 can represent the true phase difference between the two signals X1 and X2 in Table 1 at sampling point 2, and so on.

[0063] The header (first row and first column) in Table 2 is for explanation only. The label Y of each sample in the actual model data set does not include the header, and "Data 1" to "Data 2000" in Table 2 are only used as examples. The actual data shall be based on the collected signal.

[0064] Table 2

[0065] Sampling point number The phase difference between X1 and X2 at the corresponding sampling point 1 Data 1 2 Data 2 … … 2000 Data 2000

[0066] Optionally, the method for obtaining each sample in the model data set in this embodiment may be: using a simulation tool (such as a MATLAB tool) in the relevant technical field to perform simulation operations to obtain the above-mentioned samples. As an example, the simulation tool may be configured according to the parameters shown in Table 3 below, and then the simulation tool may be used to simulate multi-antenna reception signals under different channel conditions (specifically, different signal-to-noise ratios). The obtained model data set may be stored in the fifth version of the hierarchical data format (HierarchicalData Format 5, HDF5) or other formats. The model data set may contain any number of samples. For example, when simulating according to the parameters shown in Table 3, a total of 2.1 million samples containing 100,000*21 may be obtained, where 21 is the 21 different signal-to-noise ratio conditions shown in Table 3.

[0067] The signal type of the input signal in the sample can be any signal type corresponding to a digital transmission signal system, such as binary phase shift keying (BPSK), quadrature phase shift keying (QPSK), 8 phase shift keying (8PSK) signal, etc.

[0068] Table 3

[0069] signal conditions Parameter Value Signal-to-noise ratio / decibel (dB) [-20，0] Delay / nanosecond (ns) [10，10] Initial phase / degree (°) [-180，180] Amplitude / Volt (V) [0，10]

[0070] Optionally, a method for obtaining each sample in the model data set may also be to collect a signal received by a real antenna array and analyze the received signal to obtain a corresponding sample.

[0071] Alternatively, the above two methods may be combined, that is, a part of the samples are obtained through simulation, and a part of the samples are obtained by collecting signals from real antennas.

[0072] The phase estimation model used in this embodiment may be a convolutional neural network (CNN) model of any structure. Generally, a CNN model may include an input layer, multiple convolutional layers, an activation layer, a pooling layer, a fully connected layer, and an output layer.

[0073] The input layer is used to receive the normalized input signal. In this embodiment, the input layer can be a tensor of size (batch_size, 4, data_length), where batch_size is the preset batch size, 4 represents the number of input channels, which are recorded as channels 1 to 4 in sequence. In the samples of the model data set, the corresponding real and imaginary parts of the complex input signals X1 and X2 are exactly 4 different features, that is, channels 1 to 4 correspond to the real part of X1, the imaginary part of X1, the real part of X2, and the imaginary part of X2 in sequence. data_length is the length of each sample. Combined with the above example, if the input signal includes 2000 sampling points, data_length is equal to 2000.

[0074] The phase estimation model used in this embodiment may have multiple consecutive one-dimensional convolutional layers after the input layer, and each convolutional layer is responsible for extracting features of different levels from the input data. As some examples, Figure 1 As shown, the phase estimation model may include a total of 23 convolutional layers, wherein the 1st and 2nd convolutional layers are the first stage, the 3rd to 5th convolutional layers are the second stage, and the 6th to 23rd convolutional layers are the third stage.

[0075] In the first stage, the first convolution layer accepts input from 4 channels of the input layer and outputs data from 8 channels. The convolution kernel size is 5, the padding is 2, the stride is 1, and it is followed by a PReLU activation function. The second convolution layer in the first stage continues to receive data from 8 channels of the previous layer and also outputs data from 8 channels. The convolution kernel size is 5, the stride is 2, there is no padding, and it is followed by a PReLU activation function and a batch normalization layer.

[0076] In the second stage, the convolution layer of the first layer receives the 8-channel data output by the batch normalization layer of the first stage as input, and outputs 16-channel data. The convolution kernel size is 5, the stride is 2, and the padding is 2. The two convolution layers after the second stage continue to receive the 16-channel data of the previous convolution layer as input, and both output 16-channel data. The convolution kernel size is 5 and the stride is 2. The second convolution layer of the second stage is padded with 2, and the third convolution layer of the second stage is padded with 0. The third convolution layer of the second stage is followed by a PReLU activation function and a batch normalization layer.

[0077] In the third stage, the convolution layer of the first layer receives the 16-channel data output by the batch normalization layer of the second stage as input, and outputs 25-channel data. The convolution kernel size is kept at 5 and the padding is 2. In addition, in the third stage, the stride is set to 2 for every four convolution layers, and the PReLU activation function and batch normalization layer are set for every four convolution layers to enhance the nonlinear expression ability and training stability of the model.

[0078] The advantage of using the above 23 convolutional layers to process data is that through multi-level convolution operations and gradual dimensionality reduction, the phase estimation model can extract rich local and global features from multi-channel time series data while reducing the amount of computation and preventing overfitting.

[0079] In some embodiments, the initial weight of each convolutional layer can be determined using the Kaiming normal distribution initialization method to ensure the reasonable distribution of weights, further improve the performance of the model, set the variance of the weights to be calculated based on the number of output units, and specify the activation function as ReLU. And the weight of the batch normalization layer can be initialized to 1, and the bias term of the batch normalization layer can be initialized to 0.

[0080] The activation layer can use the parametric rectified linear activation function (Parametric ReLU, PReLU), which is a parameterized ReLU activation function that allows the slope in the negative region to be learnable. The traditional ReLU function sets all values ​​to 0 in the negative region, while PReLU has a learnable slope parameter A in the negative region. The PReLU activation function f(x) can be expressed as formula (1).

[0081]

[0082] The pooling layer is used to reduce the spatial size of the feature map, thereby reducing the number of parameters, reducing the computational complexity, and helping the model extract more advanced features. In the phase estimation model used in this embodiment, an explicit pooling layer may not be set, but the effect of the pooling layer is achieved by setting the stride of the convolutional layer.

[0083] The main function of the fully connected layer is to map the feature vector extracted by the convolutional layer to the output category or regression target. In convolutional neural networks, the convolutional layer is usually used to extract local features, while the fully connected layer is used to integrate these local features to form a global feature representation. Before entering the fully connected layer, the feature map output by the convolutional layer needs to be flattened to convert it from a multi-dimensional tensor to a one-dimensional vector.

[0084] In the phase estimation model of this embodiment, the fully connected layer can be composed of multiple linear layers and activation functions, including five fully connected layers in total. The input of the first fully connected layer can be a 200-dimensional feature vector, and the output can be a 100-dimensional feature vector; the input of the second fully connected layer can be a 100-dimensional feature vector, and the output can be a 50-dimensional feature vector; the input of the third fully connected layer can be a 50-dimensional feature vector, and the output can be a 25-dimensional feature vector; the input of the fourth fully connected layer can be a 25-dimensional feature vector, and the output can be a 10-dimensional feature vector; the input of the fifth fully connected layer can be a 10-dimensional feature vector, and the output can be a 1-dimensional feature vector. There is a PReLU activation function between every two fully connected layers.

[0085] In addition, the Kaiming normal distribution initialization method is used to initialize the weights of the fully connected layer. The variance of the weights can be calculated based on the number of output units, and the activation function is specified as the Rectified Linear Unit (ReLU).

[0086] It can be seen that the phase estimation model of this embodiment has a large depth, including 23 convolutional layers and 5 fully connected layers, etc., which enables the phase estimation model to learn more complex patterns.

[0087] The structures of the various layers in the above CNN model are only taken as examples. In other optional embodiments, the various layers contained in the phase estimation model may also have other structures different from the above structures, and are not limited to the above structures.

[0088] Before using the model data set to train the above phase estimation model, the samples in the model data set may be preprocessed to obtain preprocessed samples. The preprocessing process may include:

[0089] 1. Data normalization, that is, normalizing the input signal and label data (such as the data in Table 1 and Table 2) to the range of 0 to 1 to ensure the consistency of the data input to the model.

[0090] 2. Data partitioning, that is, dividing the model data set into training set and validation set. Specifically, it can be divided according to a ratio of 9:1, that is, 90% of the data is used as the training set and the remaining 10% of the data is used as the validation set. It can also be divided according to other ratios, and random indexing is used during division to ensure the randomness and representativeness of the data.

[0091] 3. Format conversion, that is, converting the format of the input data into a tensor format that matches the input layer.

[0092] During training, the model to be trained after initialization can be loaded onto a graphics processing unit (GPU) for training. In this embodiment, the mean square error loss function MSELoss can be used to calculate the predicted phase difference and the actual phase difference (i.e., the phase difference contained in the label) output by the phase estimation model to obtain the model loss; an adaptive learning rate optimization algorithm is used, the initial value of the learning rate is 0.01, and the cosine annealing learning rate scheduling strategy is used to update the learning rate during training. The upper limit of the number of iterations (equivalent to the number of times the entire training data set is traversed during the training process) epochs can be 50, and the minimum learning rate is 1*10 -5 The training parameters can be set to the values ​​in Table 4, or to other values ​​as needed.

[0093] The definition of the mean square error loss function and the method of using the function to calculate the predicted phase difference and the actual phase difference to obtain the model loss can be found in the relevant prior art and will not be described in detail.

[0094] Table 4

[0095] parameter value The number of times the entire training data set is traversed during training (epochs) 50 The number of samples processed by the model in one iteration (batch_size) 128 Optimizer that updates model parameters Adam Learning Rate Scheduler Cosine annealing Initial value of learning rate 0.01 Minimum learning rate <![CDATA[1*10 -5 ]]> Loss Function MSELoss Hardware used for training 1 A100 GPU

[0096] The training process may include:

[0097] 1. Load two sets of input signals for each sample after preprocessing from the training set in batches.

[0098] 2. Forward propagation of the input signal through the model to be trained, specifically including: the input layer converts the dimension of the input signal to obtain data in the input format suitable for the convolution layer, and then the data of the input layer is processed through multiple convolution layers, batch normalization layers and activation functions to extract the features of the input signal. The features of the convolution layer input are flattened into a one-dimensional vector in the activation layer for input to the fully connected layer. The fully connected layer obtains the one-dimensional vector output by the activation layer, and maps the one-dimensional vector to the final prediction result through multiple linear layers and activation functions, that is, the predicted phase difference between the two sets of input signals. The model to be trained has Figure 1 For the model of the structure shown, before training begins, the values ​​of the weight parameters in the model can be determined by the aforementioned initialization method.

[0099] Steps 1 and 2 are equivalent to: processing the samples with the model to be trained to obtain the predicted phase difference corresponding to the samples.

[0100] 3. Calculate the predicted phase difference and the actual phase difference based on the loss function to obtain the model loss.

[0101] 4. Perform back propagation based on model loss, including clearing the gradient calculated in the previous back propagation, recalculating the gradient based on the current model loss, and updating the model parameters based on the calculated gradient. The updated model continues to process the input signals of subsequent samples until the training end conditions are met.

[0102] The training end condition may be that the number of traversals of the samples in the training set reaches the set upper limit of the number of iterations, that is, epochs in Table 4. For example, if epochs is equal to 50, then after the training process processes all the samples in the training set, the number of traversals increases by 1, and then each sample in the training set is processed again according to the training process until the number of traversals is equal to 50, and it is determined that the training condition is met, wherein the initial value of the number of traversals is 0.

[0103] It can be understood that each time the model to be trained processes a sample, a model loss corresponding to the sample can be obtained according to step 3, and back propagation can be performed according to the model loss corresponding to the sample.

[0104] 5. Record the average loss on the training set. That is, at the end of each epoch, calculate the average value of the model loss obtained when the model to be trained processes all samples in the training set, and use this average value as the total loss of the model on the training set at the end of the epoch.

[0105] 6. Set the checkpoint save point. At the end of each epoch, save the state dictionary, loss, best loss, and optimizer state of the model to be trained; and set the best model to be saved. That is, at the end of each epoch, if the total loss of the current model on the training set is less than the historical best loss, save the current model as the best model.

[0106] When the training conditions are met and the above training process is ended, the best model at this time can be used as the phase estimation model used in this embodiment.

[0107] The historical best loss is the minimum value of the total loss corresponding to each previous epoch except the current epoch.

[0108] At the end of each epoch, a state dictionary can be saved. The state dictionary may include the values ​​of each weight in the model to be trained at the corresponding moment. Based on a state dictionary, the model at the corresponding moment can be restored.

[0109] An epoch can be understood as the process of traversing and processing all samples in the training set in one go. For example, assuming that the training set has 1.89 million samples, an epoch is considered to have ended when these 1.89 million training samples are processed according to the above training process.

[0110] Steps 3 to 6 are equivalent to: training the model to be trained according to the predicted phase difference and the actual phase difference of the sample until the training conditions are met to obtain a phase estimation model.

[0111] The total number of parameters of the optimal model finally trained can be 85,300.

[0112] After obtaining the above phase estimation model, the phase estimation model can be verified using a verification set. The verification process can be as follows:

[0113] 1. Load two sets of input signals for each sample in batches after preprocessing from the validation set.

[0114] 2. The input signal is forward propagated through the phase estimation model to obtain the phase difference between the two sets of input signals. The processing process can be referred to step 2 of the aforementioned training process and will not be repeated here.

[0115] 3. Calculate the predicted phase difference and the actual phase difference based on the loss function to obtain the model loss.

[0116] 4. Record the average loss on the validation set, that is, the average value of the model loss obtained when the phase estimation model processes each sample in the validation set.

[0117] 5. Save the predicted phase difference and the actual phase difference for subsequent analysis.

[0118] The above verification process can also be iterated multiple times. Each iteration traverses all samples in the verification set, and the average loss corresponding to all samples in the verification process is recorded as the total loss of this iteration. When the number of iterations reaches the upper limit of the number of iterations, that is, epochs in Table 4, the verification process ends.

[0119] The change curves of the model loss value during training and the model loss value during verification can be seen in Figure 2 .

[0120] The comparison between the predicted phase difference and the actual phase difference of the trained phase estimation model on the validation set can be seen in Figure 3 .

[0121] It can be seen from the above schematic diagram that the phase estimation model trained according to the above method can accurately predict the phase difference between two sets of input signals.

[0122] After the phase estimation model is obtained in the above manner, the phase estimation model can be used to Figure 4 The CNN-based high-precision delay estimation method for broadband digital transmission signals shown in the figure estimates the delay of signals received by N antennas. The method may include the following steps.

[0123] S401, obtaining multiple groups of target signals, each group of target signals is received by one antenna, and the target signals are discrete signals in complex time domain form.

[0124] The form of each group of target signals can refer to the form of signal X1 or X2 in Table 1, that is, a group of target signals can include multiple sampling points in complex form, each sampling point includes a real part data and an imaginary part data, and a sequence composed of multiple continuous sampling points is equivalent to a group of discrete signals in complex time domain form.

[0125] In this step, the intermediate frequency signal received by each antenna may be sampled to obtain a group of target signals corresponding to the antenna.

[0126] These target signals may be broadband digital transmission signals for which time delay difference and initial phase difference need to be determined.

[0127] Optionally, the trained phase estimation model used in this embodiment can be integrated into a communication receiver software platform. Therefore, the executor of the method of this embodiment can be a communication receiver software platform or a communication receiver that integrates the above-mentioned phase estimation model.

[0128] S402, performing channel simulation on each group of target signals to obtain simulated target signals with phase errors and noise added thereto.

[0129] The method of performing channel simulation on the target signal can be referred to in the relevant prior art and will not be described in detail.

[0130] Optionally, the phase error and noise added to each group of target signals may be different.

[0131] For example, assuming that N groups of target signals are obtained in S401, which are received by antenna 1 to antenna N in sequence, the phase error added to the target signal of antenna 1 in S402 can be recorded as e iφ1 , the added noise can be recorded as n1, and the phase error added to the target signal of antenna 2 can be recorded as e iφ2 , the added noise can be recorded as n2, and so on, the phase error added to the target signal of antenna n can be recorded as e iφN , the added noise can be recorded as n N , where i represents a complex unit, φ1 to φ N are preset and different parameters, n1 to n N are preset and different parameters, and e is the base of the natural logarithm.

[0132] S403, processing the reference signal and each simulated target signal except the reference signal according to the phase estimation model to obtain a predicted phase difference of each simulated target signal except the reference signal relative to the reference signal, wherein the reference signal is any one of multiple groups of simulated target signals, and the phase estimation model is a convolutional neural network model.

[0133] In S403, for each group of simulated target signals, the group of simulated target signals and the reference signal can be input into the phase estimation model, and forward propagation is performed in the phase estimation model according to step 2 of the aforementioned training process, and finally the predicted phase difference of the simulated target signal output by the phase estimation model relative to the reference signal is obtained.

[0134] As some examples, the simulated target signal received by antenna 1 may be used as a reference signal.

[0135] In combination with the above example, the phase estimation model can be used to process the simulated target signal of antenna 1 and the simulated target signal of antenna 2 to obtain the predicted phase difference of the simulated target signal of antenna 2 relative to the simulated target signal of antenna 1. The phase estimation model can be used to process the simulated target signal of antenna 1 and the simulated target signal of antenna 3 to obtain the predicted phase difference of the simulated target signal of antenna 3 relative to the simulated target signal of antenna 1. By analogy, the predicted phase difference of the simulated target signals from antenna 2 to antenna N relative to the simulated target signal can be obtained.

[0136] S404, determining a time delay difference and an initial phase difference of each simulated target signal except the reference signal relative to the reference signal according to a predicted phase difference corresponding to each simulated target signal except the reference signal.

[0137] In step S404, the predicted phase difference corresponding to each group of target signals except the reference signal may be input into a linear fitter to obtain the delay difference and initial phase difference corresponding to each group of target signals except the reference signal.

[0138] The specific working principle of the linear fitter can be found in the relevant prior art and will not be described in detail.

[0139] In some optional embodiments, the method of this embodiment may further include the following steps:

[0140] Obtaining a time delay difference and an initial phase difference of a reference signal relative to a reference signal;

[0141] The performance of the phase estimation model is detected based on the time delay difference and initial phase difference of the reference signal relative to the reference signal.

[0142] In this embodiment, a group of reference signals can be copied to obtain two identical groups of reference signals, and the two groups of reference signals can be input into a phase estimation model to obtain a predicted phase difference between the reference signal output by the phase estimation model and the reference signal, and the predicted phase difference between the reference signal and the reference signal can also be input into a linear fitter to obtain a time delay difference and an initial phase difference between the reference signal and the reference signal.

[0143] The method for detecting the performance of the phase estimation model according to the time delay difference and initial phase difference of the reference signal relative to the reference signal may be:

[0144] When the time delay difference and the initial phase difference of the reference signal relative to the reference signal are both equal to 0, it is determined that the performance of the phase estimation model is normal;

[0145] When at least one of the time delay difference and the initial phase difference of the reference signal relative to the reference signal is not equal to 0, it is determined that the performance of the phase estimation model is abnormal.

[0146] That is to say, the performance of the model can be detected by judging whether the delay difference and the initial phase difference of the reference signal relative to the reference signal are equal to 0. If the delay difference and the initial phase difference are equal to 0, it means that the performance of the phase estimation model is normal at this time. If at least one of the delay difference and the initial phase difference is not equal to 0, it means that the performance of the phase estimation model is abnormal at this time. At this time, the corresponding alarm prompt can be output to prompt relevant personnel to optimize the phase estimation model in time, thereby evaluating and feedbacking the performance of the phase estimation model in real time, discovering the deviation of the model in time and giving an alarm, and ensuring real-time monitoring of the performance of the phase estimation model.

[0147] Optionally, the model performance can be determined to be normal when the delay difference and the initial phase difference of the reference signal relative to the reference signal are close to 0, that is, the absolute value of the delay difference and the absolute value of the initial phase difference are both less than or equal to a specific threshold, and the model performance can be determined to be abnormal when the delay difference and the initial phase difference of the reference signal relative to the reference signal are not close to 0, that is, at least one of the absolute value of the delay difference and the absolute value of the initial phase difference is greater than a specific threshold.

[0148] Optionally, when the phase estimation model is used to process the reference signal and the target signal, the reference signal and the target signal may also be preprocessed according to the aforementioned method for preprocessing the input signal of the sample.

[0149] The beneficial effects of this embodiment are:

[0150] A high-precision delay estimation method for broadband digital signals based on CNN is proposed. The broadband digital signals are trained using CNN, and the delay difference and initial phase difference of the broadband digital signals are determined using the trained model.

[0151] Compared with traditional methods, the CNN model can maintain high estimation accuracy under different signal-to-noise ratio conditions and does not require cross-correlation operations between signals, which greatly reduces the amount of phase difference estimation calculations and enhances real-time performance and efficiency, thereby improving the accuracy and efficiency of delay estimation, and further improving the overall performance of the antenna array system.

[0152] Furthermore, using the CNN model to estimate the signal phase difference between antennas can automatically learn features from the data, reducing the need for manual feature extraction and simplifying the entire processing flow;

[0153] In addition, since the CNN model is adopted, the method is scalable, including scale expansion and function expansion. The scale expansion includes that the CNN model can be easily expanded to increase its scale by increasing the number of layers or nodes; the function expansion includes the subsequent expansion while processing phase difference estimation, and jointly optimizing other related tasks (such as channel estimation, signal classification, etc.), thereby improving model performance to adapt to larger-scale data processing needs, and the method of this embodiment supports implementation on a distributed computing framework, so it can facilitate training and inference on large-scale clusters.

[0154] Furthermore, the method of this embodiment can also detect the performance of the model used in real time, promptly discover the deviation of the model and issue an alarm, ensuring real-time monitoring of the CNN model performance.

[0155] Based on the above method, this embodiment also provides a high-precision delay estimation system for broadband digital transmission signals based on CNN, see Figure 5 The system may include the following modules.

[0156] The information source module 501 is used to obtain multiple groups of target signals, each group of target signals is received by one antenna, and the target signals are discrete signals in complex time domain form;

[0157] A simulation module 502, configured to perform channel simulation on each group of target signals to obtain a simulated target signal with phase error and noise added;

[0158] an estimation module 503, configured to process the reference signal and each simulated target signal other than the reference signal according to a phase estimation model to obtain a predicted phase difference of each simulated target signal other than the reference signal relative to the reference signal, wherein the reference signal is any one of the multiple groups of simulated target signals, and the phase estimation model is a convolutional neural network model;

[0159] The determination module 504 is used to determine the delay difference and initial phase difference of each simulated target signal except the reference signal relative to the reference signal according to the predicted phase difference corresponding to each simulated target signal except the reference signal.

[0160] Optionally, the determination module 504 is further configured to:

[0161] Obtaining a time delay difference and an initial phase difference of a reference signal relative to a reference signal;

[0162] The performance of the phase estimation model is detected based on the time delay difference and initial phase difference of the reference signal relative to the reference signal.

[0163] Optionally, when the determination module 504 detects the performance of the phase estimation model according to the delay difference and the initial phase difference of the reference signal relative to the reference signal, it can be used to:

[0164] When the time delay difference and the initial phase difference of the reference signal relative to the reference signal are both equal to 0, it is determined that the performance of the phase estimation model is normal;

[0165] When at least one of the time delay difference and the initial phase difference of the reference signal relative to the reference signal is not equal to 0, it is determined that the performance of the phase estimation model is abnormal.

[0166] Optionally, a training module 505 is also included, which is used to:

[0167] Processing samples with the model to be trained to obtain predicted phase differences corresponding to the samples, where the number of samples is multiple, and each sample includes two sets of input signals and a set of labels;

[0168] Each set of input signals of the sample is a discrete signal in the complex time domain form, and the label of the sample includes label data corresponding to the sampling points of the input signal one by one, and the label data is the real phase difference between the two sets of input signals of the sample at the corresponding sampling points;

[0169] The model to be trained is trained according to the predicted phase difference and the actual phase difference of the sample until the training condition is met to obtain a phase estimation model.

[0170] Optionally, the training module 505 is further used for:

[0171] Preprocessing the sample to obtain a preprocessed sample;

[0172] Process the sample with the model to be trained, including:

[0173] Process the preprocessed samples with the model to be trained;

[0174] The preprocessing includes at least one of data normalization, data division and format conversion.

[0175] The working principle of the high-precision time delay estimation system for broadband digital transmission signals based on CNN in this embodiment can be found in the relevant steps of the high-precision time delay estimation method for broadband digital transmission signals based on CNN in any of the aforementioned embodiments, and will not be repeated here.

[0176] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0177] For the convenience of description, the above system or device is described by dividing it into various modules or units according to its functions. Of course, when implementing the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0178] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0179] Finally, it should be noted that, in this article, relational terms such as first, second, third and fourth are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0180] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A high-precision delay estimation method for broadband digital transmission signals based on convolutional neural network (CNN), characterized in that: include: Acquire multiple groups of target signals, each group of target signals is received by one antenna, and the target signals are discrete signals in complex time domain form; Performing channel simulation on each group of the target signals to obtain simulated target signals with added phase error and noise; Processing a reference signal and each of the simulated target signals except the reference signal according to a phase estimation model to obtain a predicted phase difference of each of the simulated target signals except the reference signal relative to the reference signal, wherein the reference signal is any one of a plurality of groups of simulated target signals, and the phase estimation model is a convolutional neural network model; wherein the method for obtaining the phase estimation model comprises: processing samples with a model to be trained to obtain a predicted phase difference corresponding to the samples, wherein the number of the samples is multiple, and each of the samples comprises two groups of input signals and a group of labels; wherein each group of the input signals of the samples is a discrete signal in a complex time domain form, and the label of the sample comprises label data corresponding one-to-one to the sampling points of the input signal, and the label data is the real phase difference between the two groups of input signals of the sample at the corresponding sampling points; training the model to be trained according to the predicted phase difference and the real phase difference of the samples until the training conditions are met to obtain the phase estimation model; According to the predicted phase difference corresponding to each of the simulated target signals except the reference signal, the time delay difference and initial phase difference of each of the simulated target signals except the reference signal relative to the reference signal are determined.

2. The method according to claim 1, characterized in that Also includes: Copying a group of the reference signal to obtain two identical groups of reference signals, inputting the two groups of reference signals into the phase estimation model to obtain a predicted phase difference of the reference signal output by the phase estimation model relative to the reference signal, inputting the predicted phase difference of the reference signal relative to the reference signal into a linear fitter to obtain a time delay difference and an initial phase difference of the reference signal relative to the reference signal; The performance of the phase estimation model is detected according to the time delay difference and the initial phase difference of the reference signal relative to the reference signal.

3. The method according to claim 2, characterized in that The detecting the performance of the phase estimation model according to the delay difference and the initial phase difference of the reference signal relative to the reference signal comprises: When the delay difference and the initial phase difference of the reference signal relative to the reference signal are both equal to 0, determining that the performance of the phase estimation model is normal; When at least one of the time delay difference and the initial phase difference of the reference signal relative to the reference signal is not equal to 0, it is determined that the performance of the phase estimation model is abnormal.

4. The method according to claim 1, characterized in that Before processing the sample with the model to be trained, the method further includes: Preprocessing the sample to obtain a preprocessed sample; The processing of samples using the model to be trained includes: Processing the preprocessed samples with the model to be trained; The preprocessing includes at least one of data normalization, data division and format conversion.

5. A high-precision delay estimation system for broadband digital transmission signals based on convolutional neural network (CNN), characterized in that: include: A signal source module, used to obtain multiple groups of target signals, each group of target signals is received by one antenna, and the target signals are discrete signals in complex time domain form; A simulation module, used for performing channel simulation on each group of target signals to obtain simulated target signals with added phase error and noise; an estimation module, configured to process a reference signal and each of the simulated target signals except the reference signal according to a phase estimation model to obtain a predicted phase difference of each of the simulated target signals except the reference signal relative to the reference signal, wherein the reference signal is any one of a plurality of groups of the simulated target signals, and the phase estimation model is a convolutional neural network model; A determination module, configured to determine a time delay difference and an initial phase difference of each of the simulated target signals other than the reference signal relative to the reference signal according to a predicted phase difference corresponding to each of the simulated target signals other than the reference signal; A training module, used to: process samples with a model to be trained to obtain predicted phase differences corresponding to the samples, wherein the number of the samples is multiple, and each of the samples includes two groups of input signals and a group of labels; wherein each group of the input signals of the samples is a discrete signal in a complex time domain form, and the labels of the samples include label data corresponding to the sampling points of the input signals one by one, and the label data is the real phase difference between the two groups of input signals of the samples at the corresponding sampling points; The model to be trained is trained according to the predicted phase difference and the true phase difference of the sample until a training condition is met, so as to obtain the phase estimation model.

6. The system according to claim 5, characterized in that The determining module is also used for: Copying a group of the reference signal to obtain two identical groups of reference signals, inputting the two groups of reference signals into the phase estimation model to obtain a predicted phase difference of the reference signal output by the phase estimation model relative to the reference signal, inputting the predicted phase difference of the reference signal relative to the reference signal into a linear fitter to obtain a time delay difference and an initial phase difference of the reference signal relative to the reference signal; The performance of the phase estimation model is detected according to the time delay difference and the initial phase difference of the reference signal relative to the reference signal.

7. The system according to claim 6, characterized in that When the determination module detects the performance of the phase estimation model according to the delay difference and the initial phase difference of the reference signal relative to the reference signal, it is used to: When the delay difference and the initial phase difference of the reference signal relative to the reference signal are both equal to 0, determining that the performance of the phase estimation model is normal; When at least one of the time delay difference and the initial phase difference of the reference signal relative to the reference signal is not equal to 0, it is determined that the performance of the phase estimation model is abnormal.

8. The system according to claim 5, characterized in that The training module is also used to: Preprocessing the sample to obtain a preprocessed sample; The processing of samples using the model to be trained includes: Processing the preprocessed samples with the model to be trained; The preprocessing includes at least one of data normalization, data division and format conversion.

Images