A deep learning-based underwater acoustic PN signal time-domain coarse acquisition method and system

Abstract

The application belongs to the technical field of synchronous signal capture, and discloses a kind of underwater acoustic PN signal time domain coarse capture method and system based on deep learning, adopt receiving end local carrier compression algorithm to convert PN correlation gain amplitude frequency deviation response into the approximate flat of multiple SINC function module non-coherent accumulation, based on the circular convolution of local carrier compression time domain parallel correlation result output is used for TCN-Seq2Seq model provides the correlation waveform feature of large doppler tolerance;Utilize TCN-Seq2Seq model to realize the time domain synchronization of low signal-to-noise ratio condition based on the circular correlation output waveform of carrier compression, identify and analyze the received waveform, output waveform time domain range, obtain the coarse estimation of Doppler frequency deviation, complete time domain frequency domain coarse capture.The application greatly reduces the computing overhead, and has high value for the application of underwater acoustic communication of PN signal in military and biological friendly direction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of synchronous signal capture, and particularly relates to a time-domain coarse capture method for underwater acoustic PN signals based on deep learning. BACKGROUND

[0002] At present, a communication frame structure is generally composed of a synchronous signal and an information main body. A receiving end completes phase synchronization and carrier synchronization by capturing the synchronous signal, thereby laying a foundation for demodulation and decoding of the information main body part. Synchronous signal capture failure may lead to serious packet loss, and it is difficult to correct errors through channel coding. Therefore, synchronous head capture is a basic prerequisite for underwater acoustic communication. Pseudo-Noise (PN) synchronous signals are widely used in biological-friendly and military underwater acoustic communication application scenarios due to their good anti-multipath ability, low power spectrum density, and the ability to carry out channel estimation. PN synchronous signals are extremely sensitive to Doppler. Traditional PN signal capture technologies mainly include a serial search method of time-frequency two-dimensional search, a PMF-FFT frequency domain parallel search method, a circular convolution time domain parallel search method, and the like. In recent years, time-frequency compression capture algorithms, folding (XFAST) or double folding capture algorithms, and the like are all improvements or fusions of the above three basic capture algorithms. The serial search method was first proposed. It uses a certain step size to perform grid search in the time domain and the frequency domain. The idea is direct, but the calculation overhead is large, and it is not suitable for underwater communication nodes with limited energy. The PMF-FFT realizes synchronous capture through the FFT peak value of the local correlation value. The time domain search step is generally set to half of the chip sample. The search speed is slow. The circular convolution time domain parallel search realizes correlation operation through frequency domain multiplication and realizes long time domain signal parallel search through zero padding. In recent years, this technology and its improved algorithms have become a research hotspot. Related research mainly aims to reduce the calculation complexity and improve the capture speed. However, the Doppler in underwater acoustic communication is 10 times that of radio frequency communication. The search range of the fixed step Doppler frequency offset is greatly expanded. There is a large time delay in the application scenario of underwater acoustic PN signal capture and synchronization for underwater mobile platforms. At the same time, the calculation overhead of the underwater node is large, which significantly reduces the service life of the underwater node. Inspired by the good performance of deep learning in the voice field, using deep learning to solve problems in underwater acoustic communication has gradually become a new research direction, mainly involving underwater acoustic channel estimation and equalization, receiving decoding, adaptive modulation, signal blind detection estimation, and the like. However, there is still a lack of application research on deep learning in underwater acoustic PN synchronous signal capture. 3 BACKGROUND

[0002] At present, a communication frame structure is generally composed of a synchronous signal and an information main body. A receiving end completes phase synchronization and carrier synchronization by capturing the synchronous signal, thereby laying a foundation for demodulation and decoding of the information main body part. Synchronous signal capture failure may lead to serious packet loss, and it is difficult to correct errors through channel coding. Therefore, synchronous head capture is a basic prerequisite for underwater acoustic communication. Pseudo-Noise (PN) synchronous signals are widely used in biological-friendly and military underwater acoustic communication application scenarios due to their good anti-multipath ability, low power spectrum density, and the ability to carry out channel estimation. PN synchronous signals are extremely sensitive to Doppler. Traditional PN signal capture technologies mainly include a serial search method of time-frequency two-dimensional search, a PMF-FFT frequency domain parallel search method, a circular convolution time domain parallel search method, and the like. In recent years, time-frequency compression capture algorithms, folding (XFAST) or double folding capture algorithms, and the like are all improvements or fusions of the above three basic capture algorithms. The serial search method was first proposed. It uses a certain step size to perform grid search in the time domain and the frequency domain. The idea is direct, but the calculation overhead is large, and it is not suitable for underwater communication nodes with limited energy. The PMF-FFT realizes synchronous capture through the FFT peak value of the local correlation value. The time domain search step is generally set to half of the chip sample. The search speed is slow. The circular convolution time domain parallel search realizes correlation operation through frequency domain multiplication and realizes long time domain signal parallel search through zero padding. In recent years, this technology and its improved algorithms have become a research hotspot. Related research mainly aims to reduce the calculation complexity and improve the capture speed. However, the Doppler in underwater acoustic communication is 10 times that of radio frequency communication. The search range of the fixed step Doppler frequency offset is greatly expanded. There is a large time delay in the application scenario of underwater acoustic PN signal capture and synchronization for underwater mobile platforms. At the same time, the calculation overhead of the underwater node is large, which significantly reduces the service life of the underwater node. Inspired by the good performance of deep learning in the voice field, using deep learning to solve problems in underwater acoustic communication has gradually become a new research direction, mainly involving underwater acoustic channel estimation and equalization, receiving decoding, adaptive modulation, signal blind detection estimation, and the like. However, there is still a lack of application research on deep learning in underwater acoustic PN synchronous signal capture.

[0003] The prior art scheme is based on a traditional circular convolution time domain parallel search algorithm of frequency domain cyclic shift, the serial search in the frequency domain of the traditional circular convolution time domain parallel search algorithm is completed in the down-conversion carrier stripping stage, that is, a plurality of local oscillators with different frequency points of interval frequency sweeping step are introduced in the mixing process, the algorithm uses the frequency domain cyclic shift developed in recent years to replace the multi-oscillator mixing, and is more advanced than the previous algorithm in reducing the calculation overhead of the traditional algorithm. , the sending signal amplitude , the pulse shaping filter is g, is the PN code length to be captured, and the single chip duration , the PN code corresponds to a chip sequence , and the transmitting end sending passband signal corresponding to the PN code can be represented as:

[0004]

[0005] The channel impulse response can be represented as:

[0006]

[0007] Here, L is the number of received eigenrays received by the receiving node, and respectively represent the amplitude and relative time delay of the m-th eigenray to the receiving node. In this way, the received signal after the underwater acoustic multipath channel can be represented as:

[0008] , wherein

[0009] is a convolution, is additive noise. After the received signal is mixed with the local carrier oscillator and low-pass filtered, down-conversion, that is, carrier stripping, is completed, and the baseband signal is obtained. The principle of circular convolution time domain parallel search based on frequency domain cyclic shift is relatively simple, the FFT transform is used to realize time domain convolution in the frequency domain, the local code zero operation is used to realize time domain parallel search, and the frequency offset is realized by frequency domain cyclic shift, thereby reducing the calculation overhead of multi-mixing operation. This method is widely used in deep space communication PN signal acquisition. However, due to the limitation of underwater acoustic propagation speed, the Doppler factor caused by the relative motion of the platform is much larger than that of deep space communication. The present application finds that the circular convolution time domain parallel search method based on frequency domain cyclic shift is no longer suitable for PN code acquisition in the background of large Doppler underwater acoustic communication.

[0010] After the received signal is mixed with the local oscillator and low-pass filtered, the following signal is obtained:

[0011]

[0012] ​

[0013] Here is the signal compression spreading term, is the carrier frequency offset affecting term, the influence of the latter term can be offset or mitigated by frequency domain cyclic shift, however, when the PN code is longer and the Doppler factor is larger, the time domain compression spreading is no longer an approximately negligible term. When the frequency offset is small, after the PN signal is synchronized, the gain obtained by the circular convolution of the low-pass filter output signal and the local PN signal is significant, and when the gain exceeds the threshold, it is considered that the time domain synchronization is achieved.

[0014] Although the traditional circular convolution time domain parallel search algorithm based on frequency domain cyclic shift can reduce the FFT operation at M frequency offset points of the traditional circular convolution to once FFT by frequency domain cyclic shift, it still needs M times of IFFT operation. In the large Doppler process corresponding to the underwater mobile platform, the computational complexity is still large, which significantly affects the service life of the node.

[0015] Although there are many researches trying to achieve time domain and frequency domain double parallel search, due to the limitations of traditional algorithms, it has not been truly realized yet, so the large computational complexity of PN signal acquisition is still a factor hindering the military application of PN signal in underwater communication.

[0016] Through the above analysis, the problems and defects of the prior art are: 1. The influence of large Doppler factor: in underwater environment, due to the limitation of underwater sound propagation speed, the Doppler factor caused by relative motion is much larger than that in deep space communication. This leads to the decrease of PN code acquisition efficiency of circular convolution time domain parallel search method based on frequency domain cyclic shift in the background of large Doppler underwater communication.

[0017] 2. PN code length and Doppler factor: when the PN code is longer and the Doppler factor is larger, the time domain compression spreading is no longer an approximately negligible term. This means that under this condition, even if the frequency offset is small, after the PN signal is synchronized, the gain obtained by the circular convolution of the low-pass filter output signal and the local PN signal may not be enough to exceed the threshold, so that the time domain synchronization cannot be achieved.

[0018] 3. Large computational complexity: although the circular convolution time domain parallel search algorithm based on frequency domain cyclic shift can reduce the FFT operation at M frequency offset points of the traditional circular convolution to once FFT, it still needs M times of IFFT operation. Especially in the large Doppler process corresponding to the underwater mobile platform, the computational complexity is still large, which significantly affects the service life of the node.

[0019] 4. Implementation of time-frequency dual parallel search: Although there are many studies trying to implement time-frequency dual parallel search, due to the limitations of traditional algorithms, it has not been truly implemented. This results in the computational overhead of PN signal acquisition still being a factor hindering the military application of PN signals in underwater acoustic communication.

[0020] 5. Long search delay: Due to the low efficiency of traditional underwater PN synchronization signal acquisition algorithms in time or frequency domain parallel search, this leads to a long search delay, affecting real-time performance and efficiency.

[0021] Therefore, the technical problems to be solved urgently include how to improve the existing algorithm to solve the influence of large Doppler factor, how to reduce the computational overhead, how to implement time-frequency dual parallel search, and how to reduce the search delay. The solution to these problems will have a significant promoting effect on the acquisition of PN signals in underwater acoustic communication. SUMMARY

[0022] In view of the problems existing in the prior art, the present application provides a deep learning-based underwater acoustic PN signal time domain coarse acquisition method.

[0023] The present application is implemented as follows: a deep learning-based underwater acoustic PN signal time domain coarse acquisition method, comprising the following steps:

[0024] First, the local carrier compression algorithm at the receiving end is used to convert the PN correlation gain amplitude frequency offset response into a plurality of SINC function module non-coherent accumulation approximation flat, and the time domain parallel correlation result based on the circular convolution of the local carrier compression is output to provide the correlation waveform features with large Doppler tolerance for the TCN-Seq2Seq model;

[0025] Second, determine the input features of the TCN-Seq2Seq model, the input features include the circular convolution correlation output based on the local carrier compression, the local DDC waveform output without frequency offset and the original received waveform; the TCN-Seq2Seq model is constructed to include 10 residual network blocks and global pooling layers, fully connected layers and softmax layers, and finally outputs a classification sequence with the same length as the input time sequence, providing an analysis basis for subsequent time-frequency coarse synchronization;

[0026] Third, complete offline training based on target sea area simulation waveform data, and obtain a TCN-Seq2Seq acquisition model suitable for the multipath channel environment of the target sea area; in the online test stage, first carry out Fine-Tune transfer learning based on the known prior information received waveform, to compensate for the differences between the simulation stage and the actual sea multipath environment and noise environment with small computational overhead;

[0027] Fourthly, the TCN-Seq2Seq capture model result is post-processed to realize time domain coarse synchronization under a low signal-to-noise ratio condition, to obtain a Doppler frequency offset coarse estimation, and to complete time domain and frequency domain coarse capture.

[0028] Further, the local carrier compression in the first step refers to generating local subcarriers in a down-conversion process at a receiving end, entering into mixing through non-coherent accumulation, and spacing the frequency offset of each subcarrier by , PN signal pulse width, and subsequent circular convolution process, that is, using frequency domain multiplication to complete circular convolution, completing correlation of the zero-padded local baseband PN signal and the down-converted signal of the compressed carrier, and finally completing time domain parallel search synchronization through decision.

[0029] Further, the received signal is represented as:

[0030]

[0031] wherein:

[0032]

[0033] A represents signal amplitude, the PN code corresponds to a chip sequence , L is the PN code length, is the chip width, is the carrier frequency, is a Doppler factor, and the local compressed carrier is represented as:

[0034]

[0035] After mixing, the following is obtained:

[0036]

[0037]

[0038] After low-pass filtering to remove high-frequency terms, the following is obtained:

[0039]

[0040] After entering the circular convolution, when the time domain synchronization of the PN signal is aligned with the local baseband PN signal, the correlation result is:

[0041]

[0042] Let , then:

[0043]

[0044] For the frequency offset Doppler factor determined by the received signal Its correlation gain is The sum of the SINC functions.

[0045] Furthermore, in the second step, PN code acquisition synchronization training is carried out from the sequence to sequence perspective. The position of the PN code sequence is random. The label for the time segment where the PN code exists is set to 1, and the label for the time segment where the PN code does not exist is set to 0. The received signal is normalized and then labeled. The training set containing the PN code is called the positive example set. The training set also includes some pure noise signals without the PN code, which are the negative example set.

[0046] Furthermore, the ratio of the time domain length of the PN code to the time domain length of pure noise in the positive example set is 2:1, the time domain truncation length of the negative example set is the same as that of the positive example set, and the ratio of the number of positive example set samples to the number of negative example set samples is 3:1. Thus, the proportion of PN code in the time domain to the proportion of pure noise non-signal in the time domain in the overall training set is 1:1, and the same proportion of sample data is used in the validation set and the test set.

[0047] Furthermore, the underwater acoustic PN signal time-domain coarse acquisition method also includes: TCN-Seq2Seq model sequence classification capability analysis, and time-domain windowing of the target PN code passband signal to 1.5. The starting position of the PN code passband signal is randomly set and recorded, and the step size during time-domain search is... During the training set setup, white noise signals were added in 1dB increments from -30dB to 0dB as the signal-to-noise ratio condition; the training set was set with... Take the Doppler factor for the step size and then... to Conduct resampling simulation of a mobile environment; extract features under each signal-to-noise ratio (SNR) and Doppler condition to generate 300 positive example training sets, set training set labels according to the starting position and passband length of each PN code, set the corresponding position of the PN code to 1, and the rest to 0; at the same time, generate 100 sets of pure noise under each SNR and Doppler condition, and generate 100 negative example training sets after feature extraction, with all negative example training set labels being 0; a total of 400 training set input signals under each SNR condition, the number of validation sets is set to 1 / 10 of the number of training sets, and the generation method is the same as that of the training sets;

[0048] The search and capture model based on the above settings has a time-domain 1.5T step search and capture capability, that is, it has a time-domain parallel search capability.

[0049] Based on the above model setup method, the sequence is trained to the sequence classification model using a PN signal with a code length of 2047 as the capture object.

[0050] Further, the underwater acoustic PN signal time domain coarse acquisition method further comprises: TCN-Seq2Seq model result post-processing and acquisition capability analysis, including

[0051] The classification judgment logic value output by the TCN-Seq2Seq model is converted into a numerical format, and the average value of the classification results of each point and the T-1 points after the point in the range of [1:0.5*T] is taken as the CFAR model input observation value.

[0052] The model output values of the sampling points outside the T-1 points after the target point in the detection window in STEP1 are replaced by the guardian cells in the traditional CFAR, and the detection is carried out after setting the false alarm rate.

[0053] When there is a CFAR decision result, it is considered that there is a target PN signal in the window, and the central position of the CFAR decision output result is taken as the predicted time domain starting point.

[0054] Another object of the present application is to provide an underwater acoustic PN signal time domain coarse acquisition system based on deep learning, which comprises:

[0055] The local carrier compression module is used for converting the PN correlation gain amplitude frequency offset response into a plurality of SINC function module non-coherent accumulation approximate flat, and outputting the time domain parallel correlation result based on the local carrier compression circular convolution, to provide the TCN-Seq2Seq model with a large Doppler tolerance correlation waveform feature.

[0056] The TCN-Seq2Seq model construction module is used for determining the input features of the TCN-Seq2Seq model, and the input features include the circular convolution correlation output based on the local carrier compression, the local DDC waveform output without frequency offset and the original received waveform. The module includes 10 residual network blocks, a global pooling layer, a full connection layer and a softmax layer, and finally outputs a classification sequence with the same length as the input time sequence.

[0057] The training module is used for offline training based on the target sea area simulation waveform data, to obtain a TCN-Seq2Seq acquisition model suitable for the target sea area multipath channel environment.

[0058] Further, the training module carries out Fine-Tune transfer learning based on the known prior information received waveform in the online test stage, to compensate for the differences between the simulation stage and the actual sea multipath environment and noise environment with small calculation overhead.

[0059] Further, the system further comprises a post-processing module for post-processing the TCN-Seq2Seq capture model result, realizing time domain coarse synchronization under a low signal-to-noise ratio condition, obtaining a Doppler frequency offset coarse estimation, and completing time domain and frequency domain coarse capture. Meanwhile, the local carrier compression module realizes local carrier compression, that is, a local subcarrier is generated in the down-conversion process at the receiving end, enters into mixing through non-coherent accumulation, the frequency offset interval of each subcarrier is PN signal pulse width, and subsequent circular convolution process is performed, that is, the frequency domain multiplication is equivalent to circular convolution, the correlation of the zero-padded local baseband PN signal and the down-converted signal after carrier compression is completed, and finally time domain parallel search synchronization is completed through decision.

[0060] In combination with the above technical solutions and the technical problems solved, the technical solutions to be protected by the application have the following advantages and positive effects:

[0061] First, in view of the problems of low efficiency, large calculation overhead, and long search time delay of the traditional water acoustic PN synchronization signal capture algorithm in time domain or frequency domain single domain parallel search, the application proposes a PN signal time domain coarse capture device based on a Temporal Convolutional Networks Squence-to-Squence (TCN-Seq2Seq) deep learning network model, realizes time domain and frequency domain double parallel search, and significantly improves the search efficiency. Unlike the traditional capture algorithm based on feature extraction and decision, the TCN-Seq2Seq deep learning model can still realize effective recognition and time domain positioning capture under a low signal-to-noise ratio condition with long PN code correlation waveform information as a guide, plays a time reverse mirror multipath signal compression effect, breaks through the difficulty of multipath estimation of the time reverse mirror under a low signal-to-noise ratio condition, and realizes PN code time-frequency double parallel time domain coarse capture under a large Doppler background. A carrier compression circular correlation algorithm is used to provide correlation waveform information for the deep learning model, the implementation method is simple, does not increase the calculation amount, improves the system Doppler tolerance with a certain correlation gain loss, and enables the deep learning model to have a certain frequency domain parallel search capability; a post-processing method is proposed for the model output sequence classification result to obtain the signal starting position point, and the time domain capture precision is further improved on the basis of the deep learning output classification sequence. In summary, a deep learning method-based solution is proposed for the time domain coarse capture of a PN synchronization signal in a low signal-to-noise ratio and large Doppler water acoustic communication. The solution usability is verified through simulation and sea trial data, and has engineering application prospects.

[0062] The application is based on a deep learning algorithm, and a feature extraction method based on a carrier compression algorithm is proposed. The compressed relevant waveform is used as a feature, and the time domain and frequency domain double parallel search is truly realized. Only one search is needed to realize the high Doppler PN signal capture of the underwater mobile platform under the condition of low signal-to-noise ratio, which greatly reduces the calculation cost, and has high value for the application of PN signal in military and biological friendly direction of underwater acoustic communication.

[0063] Second, as the creative evidence of the invention claim, it is also reflected in the following important aspects:

[0064] (1) The technical scheme of the application fills the domestic and foreign industry technical blank:

[0065] The application first proposes to use deep learning method to realize time-frequency domain double parallel search in the process of underwater acoustic communication PN signal capture synchronization, fills the technical blank of deep learning application in the process of underwater acoustic communication PN signal synchronization, and truly realizes time-frequency domain double parallel search, which has high scientific research value and good engineering application prospect.

[0066] (2) The technical scheme of the application solves the technical problems that people have been eager to solve but have failed:

[0067] The application solves the problem of PN signal time-frequency two-dimensional search that has plagued people for a long time. The current method only realizes one-dimensional parallel search in time domain or frequency domain, and the calculation cost is still large. The application makes full use of the high generalization recognition ability of the front-end pretreatment process and convolutional neural network to waveform compression and expansion, breaks through the original limitation, and realizes time-frequency domain double parallel search.

[0068] Fourth, the underwater acoustic PN signal time domain coarse capture system based on deep learning has made the following significant technical progress:

[0069] 1) Improve the coarse capture performance: through the deep learning technology, the system can more accurately perform time domain coarse capture. The model can more accurately identify and process signals by learning the signal characteristics in different sea environments, greatly improving the accuracy of coarse capture.

[0070] 2) Enhance the universality of the system: due to the use of deep learning method, the model can learn and adapt to different sea environments, so that the system has better generalization ability. Whether it is deep sea exploration or submarine communication, the system can provide excellent performance.

[0071] 3) Improve the capture efficiency: by using TCN-Seq2Seq model and local carrier compression, the system realizes high-precision capture while greatly reducing the calculation complexity and improving the real-time performance.

[0072] 4) Optimized Doppler frequency offset estimation: the post-processing module of the system can achieve time-domain coarse synchronization and obtain coarse Doppler frequency offset estimation under low signal-to-noise ratio conditions, which is difficult to achieve in previous systems.

[0073] 5) Provide migration learning ability: during the online test phase, the system can perform Fine-Tune migration learning to compensate for the differences between the simulation phase and the actual sea multipath environment and noise environment with small computational overhead, which enables the system to quickly adapt to new environments and improves the practicality of the system. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 is a flow chart of the underwater acoustic PN signal time domain coarse acquisition method provided by the embodiment of the application;

[0075] Figure 2 is a local carrier compression process chart provided by the embodiment of the application;

[0076] Figure 3 is a peak-to-average ratio-Doppler factor response effect chart provided by the embodiment of the application;

[0077] Figure 4 is a causal dilated convolution network schematic diagram provided by the embodiment of the application;

[0078] Figure 5 is a model framework chart provided by the embodiment of the application;

[0079] Figure 6 is a signal-to-noise ratio setting schematic diagram provided by the embodiment of the application;

[0080] Figure 7 is a sequence classification accuracy chart under different multipath backgrounds provided by the embodiment of the application;

[0081] Figure 8 is a CFAR detection effect schematic diagram provided by the embodiment of the application;

[0082] Figure 9 is a sound speed profile schematic diagram provided by the embodiment of the application;

[0083] Figure 10 is a schematic diagram of the relationship between different sound source positions and receiving hydrophone positions provided by the embodiment of the application;

[0084] Figure 11 is a detection probability effect chart provided by the embodiment of the application;

[0085] Figure 12 is a test position schematic diagram provided by the embodiment of the application;

[0086] Figure 13is a learning rate schematic diagram of a frozen front-end residual convolutional neural network provided by an embodiment of the present application;

[0087] Figure 14 is a schematic diagram of obtaining the PN signal detection of each algorithm under different signal-to-noise ratio conditions. DETAILED DESCRIPTION

[0088] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0089] As shown in Figure 1 , the water acoustic PN signal time domain coarse acquisition method provided by the embodiment of the present application comprises the following steps:

[0090] S101: The PN correlation gain amplitude frequency offset response is converted into a plurality of SINC function module non-coherent accumulation approximate flat by using a local carrier compression algorithm; the time domain parallel correlation result based on the circular convolution of the local carrier compression is outputted to provide the correlation waveform features with large Doppler tolerance for the TCN-Seq2Seq model;

[0091] S102: The input features of the TCN-Seq2Seq model are determined, the input features include the circular convolution correlation output based on the local carrier compression, the local DDC waveform output without frequency offset and the original received waveform; the TCN-Seq2Seq model is constructed to include 10 residual network blocks and global pooling layers, fully connected layers and softmax layers, and finally outputs the classification sequence with the same length as the input time sequence, which provides the analysis basis for the subsequent time domain and frequency domain coarse synchronization;

[0092] S103: The offline training based on the target sea area simulation waveform data is completed to obtain the TCN-Seq2Seq acquisition model suitable for the multipath channel environment of the target sea area; in the online test stage, the Fine-Tune transfer learning based on the known prior information received waveform is carried out to compensate for the difference between the simulation stage and the actual sea multipath environment and noise environment with small calculation overhead;

[0093] S104: The TCN-Seq2Seq acquisition model result is post-processed to realize the time domain coarse synchronization under the low signal-to-noise ratio condition, obtain the Doppler frequency offset coarse estimation and complete the time domain and frequency domain coarse acquisition.

[0094] Embodiment 1

[0095] 1. Circular convolution based on local carrier compression

[0096] The embodiment proposes a local carrier compression algorithm at the receiving end, which can convert the PN correlation gain amplitude frequency offset response into the approximate flat effect of multiple SINC function module non-coherent accumulation, and can multiply the Doppler tolerance at the cost of certain gain loss. The algorithm can be directly applied to the traditional circular convolution time domain parallel search. In the application, the time domain parallel correlation result of the circular convolution based on the local carrier compression is output to provide the correlation waveform features with large Doppler tolerance for the TCN-Seq2Seq model.

[0097] As shown in Figure 2 , the local carrier compression refers to generating local subcarriers during the down-conversion process at the receiving end, and the subcarriers are incoherently accumulated and then mixed. The frequency offset interval between the subcarriers is , and , where is the PN signal pulse width. Subsequently, the circular convolution process is performed, that is, the frequency domain multiplication is used to perform circular convolution, the zero-padded local baseband PN signal is correlated with the down-converted signal of the compressed carrier, and finally the time domain parallel search synchronization is completed through decision.

[0098] The derivation process is briefly given. The received signal can be represented as:

[0099]

[0100] wherein:

[0101]

[0102] Here, A represents the signal amplitude, the PN code corresponds to the chip sequence , L is the PN code length, is the chip width, is the carrier frequency, is the Doppler factor. The local compressed carrier can be represented as:

[0103]

[0104] After mixing, ignoring noise, the following can be obtained:

[0105]

[0106]

[0107] After low-pass filtering to remove high-frequency terms, the following is obtained:

[0108]

[0109] After entering the circular convolution, when the PN signal time domain is synchronized with the local baseband PN signal, the correlation result is:

[0110]

[0111] Record Then:

[0112]

[0113]

[0114] (11)

[0115] As can be seen from equation (11), for the frequency offset Doppler factor determined by the received signal The correlation gain is The sum of SINC functions. The simulation results are given here, and the simulation and sea test parameters of the application are consistent: =7KHz, =0.34ms, L=2047, sampling factor k=4, PN code is a chaotic spread spectrum code, the number of local carrier compression C is 1, 3, 5, 7, 9, 15, and the received signal Doppler factor is taken at equal intervals in the interval [-0.33:0.33] with an interval step , and the peak-to-average ratio-Doppler factor response effect diagram is shown in Figure 3 As can be seen, for different carrier compression strategies, there are repeated peaks consistent with the number of C, and the peak position is consistent with , verifying the reasoning result of equation (11). It should be noted that when the frequency offset is large, the peak value is not only affected by the carrier phase jump (carrier compression can reduce this effect), but also by the cumulative misalignment of the chip alignment caused by waveform deformation, so even if carrier compression is used, the peak value gain will still decrease at large Doppler frequency offset compared to small Doppler frequency offset.

[0116] As can be seen from Figure 3 , after the carrier compression strategy is determined, the peak-to-average ratio of the correlation result exists SINC function up and down fluctuation, but when the frequency offset is in the range , the lower limit of the peak-to-average ratio exists a platform effect, which can ensure the time domain parallel capture under certain signal-to-noise ratio conditions. This method is simple to implement, and the local carrier compression can be designed in advance according to the PN signal pulse width length and carrier frequency , without increasing the time domain parallel search calculation amount of circular convolution, the Doppler tolerance of PN synchronization signal capture is improved.

[0117] Due to the introduction of noise interference in the local carrier compression process, the correlation gain is seriously attenuated, and the circular convolution time domain parallel search method based on carrier compression is weak in resisting low SNR conditions. In the following, the circular convolution correlation result based on carrier compression is mainly used to provide input feature sequence for the TCN-Seq2Seq model, and the traditional circular convolution search method without frequency domain parallel search and the circular convolution parallel search method based on carrier compression are used as the detection effect comparison algorithm of the method proposed in the application.

[0118] 2. PN signal time domain coarse acquisition algorithm based on TCN-Seq2Seq deep learning model

[0119] 2.1 TCN-Seq2Seq model and parameter setting

[0120] TCN-Seq2Seq is different from ordinary one-dimensional convolutional neural network (1d-CNN), which is characterized by causal convolution and dilated convolution, has the time sequence history memory function of RNN network, and avoids the problem of too large memory requirement and gradient explosion in the process of long time sequence processing. Among them, the causal convolution ensures that the length of the input sequence and the output sequence of the TCN-Seq2Seq model remains consistent, preventing information leakage; and the dilated convolution is a hollow added in the ordinary convolution, which can exponentially expand the perception field of the model. With the help of the large depth of the network obtained by the residual network structure, TCN can further expand the perception field and deeply extract the feature relationship of longer time sequence.

[0121] Based on the previous work, it is found that TCN has strong waveform recognition ability. Here, the TCN-Seq2Seq model is used to realize time domain synchronization of the circular correlation output based on carrier compression under low SNR conditions, and on this basis, the received waveform is identified and analyzed, the waveform time domain range is output, the Doppler frequency offset is coarsely estimated, and the time domain and frequency domain coarse acquisition is completed. Figure 4 It is a causal dilated convolution network diagram, and the number of convolution kernels is 3.

[0122] The input features and model architecture of the TCN-Seq2Seq model are determined by a large number of tests. The input features include the circular convolution correlation output based on local carrier compression, the local DDC waveform output without frequency offset, and the original received waveform. The model construction includes 10 residual network blocks, global pooling layer, full connection layer, and softmax layer, and finally outputs a classification sequence with the same length as the input time sequence, which provides the basis for subsequent time domain and frequency domain coarse synchronization. The model framework diagram is shown in Figure 5 , and the model setting parameters are shown in Table 1.

[0123] Table 1 TCN-Seq2Seq model design parameter setting

[0124] TCN model parameters Parameter settings Number of residual modules 10 Convolution kernel size 5 Number of convolution kernels 256 Dropout factor 0.005 Initial learning rate 0.0002 Learning rate drop rate 0.2 Learning rate drop period 10 Learning gradient threshold 2 Minimum batch training amount 5 Maximum training rounds 15

[0125] Here from the sequence to the sequence of PN code capture synchronization training, PN code sequence position is random, for the time domain segment label set to 1, for the PN code does not exist for a period of time label set to 0, after the normalization processing of the received signal is added to the label, such as Figure 6 As shown in the schematic diagram (the signal-to-noise ratio is set to-10dB). The training set containing PN code is called positive set here, and in order to adapt to the identification of non-PN code background signal, part of the pure noise signal without PN code is also set in the training set, which corresponds to the negative set. It should be pointed out that in the model training process, the standard cross entropy loss is used as the loss function, and the classification accuracy evaluation effect is greatly affected by the proportion of the overall number of samples. For example, a single PN code is added in the 5 times PN code length background noise as the capture object, then for the positive set, the model determines that it is not a PN signal (the output judgment label is all 0) and can also achieve an accuracy of 0.8; and considering the existence of the negative set, the overall label is not a PN signal and can actually obtain a higher accuracy. Therefore, it is very important to reasonably set the proportion of the positive set and the negative set, and the duty cycle of the PN code signal in the positive set for model training. The ratio of the time domain length of the PN code in the positive set to the time domain length of the pure noise used in the present application is 2:1, the time domain window length of the negative set sample is consistent with the positive set, and the ratio of the number of positive set samples to the number of negative set samples is 3:1, so that the PN code time domain proportion in the overall training set and the non-signal time domain proportion of the pure noise are 1:1, which is beneficial to the effective training of the model. The same proportion of sample data is used in the verification set and the test set.

[0126] 2.2 TCN-Seq2Seq model sequence classification ability analysis

[0127] The target PN code passband signal (PN signal pulse width ) is expanded to 1.5 in time domain, and the starting position of the PN code passband signal is randomly set and recorded, so that the step-by-step span is when searching in time domain. In order to adapt the model to the low signal-to-noise ratio condition, the white noise signal is added in the training set setting process with 1dB as the step from-30dB to 0dB as the signal-to-noise ratio condition; in addition, in order to adapt the model to the complex Doppler environment, the training set takes the Doppler factor with as the step from to A resampling simulation mobile environment is carried out. Under each SNR condition and Doppler condition, 300 groups of positive example training sets are extracted and features are generated, the training set labels are set according to the starting position of each group of PN codes and the PN code passband length (length after Doppler resampling), the position corresponding to the PN code is set to 1, and the remaining positions are set to 0; meanwhile, 100 groups of pure noise are generated under each SNR and Doppler condition, and 100 groups of negative example training sets are generated after feature extraction, and the negative example training set labels are all 0, so there are 400 groups of training set input signals under each SNR condition. The number of verification sets is 1 / 10 of the number of training sets, and the generation mode is consistent with that of the training set.

[0128] The search and capture model based on the above settings has a time domain 1.5T step search and capture capability, that is, it has a time domain parallel search capability, and based on the carrier compression and extension circular correlation Doppler tolerance, time domain and frequency domain double parallel search is realized, so the TCN-Seq2Seq model proposed in the application is a deep learning algorithm with time domain and frequency domain double parallel search capability.

[0129] According to the above model setting method, the PN signal with a code length of 2047 is taken as the training sequence of the sequence-to-sequence classification model. Here, in order to investigate the influence of multipath resources on the TCN-Seq2Seq model, the sequence classification accuracy under different multipath backgrounds is given as shown in Figure 7 . Among them, the multipath is taken from the Laoshanwan test data, different multipath data is derived from different time delay range values, channel C is a complete channel, channel B retains the first two main channel multipaths, and channel A only retains the strongest multipath. It can be seen that channel C has more rich correlation waveform features, and under the support of TCN large perception field and historical data analysis capability, it shows better classification effect under the same SNR condition.

[0130] In the traditional correlation search, under the condition that the channel impulse response cannot be accurately estimated under the condition of low SNR, the multipath signal with an amplitude less than 1 / 2 of the strongest channel is often regarded as noise and cannot contribute to the capture process. However, in the deep learning process, the correlation waveform features can be learned to realize PN signal capture under low SNR condition, and the combined contribution of multipath signals to PN signal capture can be realized without channel estimation and virtual time reflection mirror channel signal compression. The following analysis discusses channel C as the multipath background.

[0131] 2.3 TCN-Seq2Seq model result post-processing and capture capability analysis

[0132] The classification judgment sequence result output by the TCN-Seq2Seq model needs to complete the information such as existence of PN signal and PN signal time domain position, so as to complete the time domain coarse synchronization and then complete the capture.

[0133] A post-processing method based on an improved constant false alarm rate (CFAR) is proposed herein:

[0134] Step 1. Convert the classification judgment logic value output by the TCN-Seq2Seq model into a numerical format, and take the average of the classification results (logical to numerical format) of each point and the subsequent T-1 points in the range of [1:0.5*T] as the observation value of the CFAR model input.

[0135] Step 2. Replace the guardian unit in the traditional CFAR with the model output value of the sampling points outside the T-1 points after the target point in the detection window in STEP 1, and conduct detection after setting the false alarm rate.

[0136] Step 3. When there is a CFAR decision result, it is considered that there is a target PN signal in the window, and the central position of the CFAR decision output result is taken as the predicted time domain starting point.

[0137] For the problem of PN code existence judgment and starting position estimation from sequence point-by-point classification output, an improved one-dimensional constant false alarm rate (1D-CFAR) method is proposed. In this method, the decision detection unit is not a single sampling point value, but the average of a segment of data. In a 1.5*N long time domain window, the time domain starting position search judgment is carried out from the 1st point to the 0.5*N point, that is, the corresponding point and the subsequent N length are taken as the decision detection unit, the average of the N TCN-Seq2Seq model sequence classification results (0 and 1 are converted from logical values to floating point values) is taken, the guardian unit takes 10 sampling points, when the detection unit spans the both ends of the time domain window, the corresponding side guardian unit is reduced, when the detection unit is close to the both sides of the time domain window, the corresponding side guardian unit is zero. The classification results of other points in the time domain window are used as training units. According to the 1D-CFAR algorithm, the PN code detection results in the range of [1:0.5*N] under different false alarm rates can be obtained, Figure 8 The CFAR detection effect diagram according to the above method is shown in FIG. 1. It can be seen that this method can partially compensate for the errors in the output results of the TC-Seq2Seq model, and further improve the comprehensive capture ability of the algorithm.

[0138] In order to prove the creativity and technical value of the technical scheme of the present application, this part is the specific product or related technology application example of the technical scheme of the claim.

[0139] In order to verify the effectiveness of the method of the present application, a special test was carried out in Qingdao Laoshan Bay in March 2023, and the test results showed that the offline TCN-Seq2Seq model based on simulation data could realize the time-frequency domain parallel capture of PN signals under high Doppler (relative radial velocity of 6 knots and below) and low signal-to-noise ratio conditions (signal-to-noise ratio condition of-15 dB and above) after transfer learning based on limited measured signals.

[0140] BELLHOP is used to simulate the shallow water acoustic channel in the sea for offline training of the TCN-Seq2Seq model. The environmental parameters are consistent with the conditions of the Laoshan Bay test, and the sound speed profile is as shown in Figure 9 Figure 10 Different channel impulse responses are obtained according to different relationships between the sound source position and the receiving hydrophone position. The target PN signal is converted into a passband signal, and the simulated PN receiving signal waveform data of the receiving point is formed by convolving the channel and adding white noise. The de-carrier waveform data is obtained by down-conversion, and the circular correlation waveform data is obtained by carrier compression processing. The above three types of data together constitute the input features of the offline training process of the TCN-Seq2Seq model.

[0141] The Doppler factor The training background is set as a multiple Doppler factor, and the SNR training background condition is set with a step size of 1 dB. Each pair of transmitting and receiving points corresponds to a set of multipath, Doppler, and SNR conditions. The training set size is set to 400, and the offline TCN-Seq2Seq model based on the simulation channel and white noise is obtained. The test set is randomly selected from the grid position, and the random Doppler value is in the range of After post-processing, the detection probability effect of Figure 11 is obtained. In order to compare the detection effect of the TCN-Seq2Seq detector (the entire functional module of the TCN-Seq2Seq model sequence classification and subsequent post-processing process is called the TCN-Seq2Seq detector) with that of the traditional algorithm, the traditional circular correlation detection without Doppler, the traditional circular correlation detection under Doppler background, and the circular correlation detection based on carrier compression under Doppler background are given. The value range and step size of the Doppler factor are the same as those of the TCN-Seq2Seq detector. It can be seen that the traditional circular correlation detection method has no Doppler generalization ability, and although the carrier compression processing can improve the Doppler tolerance of the traditional circular correlation, the adaptability to low SNR conditions is obviously decreased. The TCN-Seq2Seq detector based on the carrier compression circular correlation waveform result can achieve a detection effect close to that of the traditional circular correlation without Doppler influence, and improves the Doppler tolerance range of the system under low SNR conditions.

[0142] 2. Transfer learning and verification based on sea trial data

[0143] A PN synchronization signal acquisition test was conducted in Laoshan Bay, China in March 2023. The test location is as shown in Figure 12 ​​The signal transmitting ship curve represents the signal transmitting ship moving path (direction is away from the receiving ship), the color change represents the radial moving speed, which is converted according to the shipborne Beidou system observation ship speed and the receiving and transmitting ship position triangular relationship, the course is first accelerated and then slowly decelerated, and the receiving and transmitting radial distance is expanded from 2.3 km to 2.9 km. The average water depth of the test sea area is 18.1 meters, and the terrain is relatively flat except for the islands, and the bottom material is muddy. During the test process, 30 frames of PN synchronous signals are transmitted, each frame contains 36 groups of PN signals for capture verification, and the time slot interval is random.

[0144] The present application directly applies the offline TCN-Seq2Seq detector obtained based on simulation data to the Laoshan Bay test data, simultaneously adopts the traditional circular convolution time domain parallel capture method and the circular convolution time domain parallel search method based on carrier compression, and the capture probability effect is as shown in Table 2. Among them, Conv-C represents the traditional circular correlation time domain search method based on carrier compression (without frequency domain serial search), Carr-C represents the circular correlation time domain search method based on carrier compression, and TCN-Seq2Seq Detector is the TCN-Seq2Seq detector obtained based on offline training of simulation data.

[0145] Table 2. Comparison of PN signal detection rates of different algorithms in sea trial

[0146] Algorithm Conv-C Carr-C TCN-Seq2Seq Detector Detection probability 0.2944 0.8111 0.8389

[0147] It can be seen that the circular convolution time domain parallel search method based on carrier compression expands the Doppler tolerance of the original circular convolution capture algorithm without frequency domain serial search, so that the PN signal detection rate is obviously improved in the analysis of sea trial data under the Doppler background. However, the detection effect of TCN-Seq2Seq Detector is far from that in the simulation stage. Here, the present application analyzes that the difference between the simulation stage and the real sea environment is the main reason. The Bellhop simulation cannot fully reflect the sea surface and sea bottom undulation, and is limited by two-dimensional simulation and model errors, so that there is a certain error between the simulation channel and the communication channel in the real sea environment. At the same time, under the relative moving background, the Doppler difference also exists between the delay eigenrays due to the grazing angle, and the sea surface undulation will add the wave period modulation to the Doppler difference, making this phenomenon more complex and difficult to reflect the real sea situation through simulation. In addition, the white noise environment used in the simulation is too ideal, and it is difficult to truly reflect the sea environment noise.

[0148] To improve the applicability of the TCN-Seq2Seq detector in real marine environments, a Fine-Tune-based transfer learning method is proposed. The last fully connected layer and classification layer of the offline TCN-Seq2Seq model based on simulation data are trained separately using real sea test data, while the learning rate of the front-end residual convolutional neural network is frozen, as shown in Figure 13 The deeper the learning layer in deep learning, the higher the degree of abstraction and the higher the relevance to the target. The front-end shallow layer only provides low-level features (feature abstraction extraction) and provides materials for the back-end classification. For marine communication nodes, only the back-end 2 layers are subjected to transfer learning, and the adaptability calculation overhead is limited, which is an algorithm attempt that can be implemented in engineering applications. Here, Fine-Tune is used to train the Fine-Tune transfer learning process for the first 10 frames of PN signal input in the test data, and the TCN-Seq2Seq-TL model after transfer learning is obtained. The output sequence classification result is processed to obtain the PN signal detection result.

[0149] To investigate the improvement of the PN signal detection capability of the deep learning model in real marine environments through transfer learning, the traditional circular convolution time-domain parallel capture, circular convolution parallel capture based on carrier compression, TCN-Seq2Seq detector without transfer learning, and TCN-Seq2Seq-TL detector trained through Fine-Tune transfer learning are applied to the last 20 frames of PN signal capture in the test data. To investigate the detection effect of each algorithm under low SNR conditions, noise is added to the test data (the SNR of the test data is set to 0 here), and the PN signal detection of each algorithm under different SNR conditions is shown in Figure 14

[0150] It can be seen that the TCN-Seq2Seq-TL detector trained through Fine-Tune transfer learning has adapted to the real marine environment conditions and basically achieved the detection effect in the simulation stage, with a significant improvement over the TCN-Seq2Seq detector without transfer learning.

[0151] In summary, the sea test results confirm the effectiveness of the method.

[0152] Example One: Deep-sea exploration application

[0153] In deep-sea exploration applications, a deep learning-based underwater acoustic PN signal time-domain coarse capture system can be used to improve the performance of underwater acoustic communication.

[0154] 1. Local carrier compression module: At the receiving end, a local subcarrier generator generates multiple local subcarriers. These local subcarriers are subjected to non-coherent accumulation and then enter the mixer, and the frequency offset interval of each subcarrier is set to the PN signal pulse width. ​

[0155] 2. TCN-Seq2Seq model construction module: In this module, a 10-layer residual network is used to construct the TCN-Seq2Seq model. The input features of the model include the circular convolution correlation output received from the local carrier compression module, the local DDC waveform output without frequency offset, and the original received waveform.

[0156] 3. Training module: offline training is performed using simulated waveform data for deep sea exploration, and a TCN-Seq2Seq capture model adapted to the deep sea multipath channel environment is obtained.

[0157] 4. Post-processing module: This module post-processes the results of the TCN-Seq2Seq capture model, realizes time domain coarse synchronization under low signal-to-noise ratio conditions, obtains coarse estimation of Doppler frequency offset, and completes time domain and frequency domain coarse capture.

[0158] Embodiment Two: Submarine Communication Application

[0159] In the submarine communication application, the deep learning-based underwater PN signal time domain coarse capture system can be used to improve the accuracy and reliability of submarine communication.

[0160] 1. Local carrier compression module: At the receiving end, a plurality of local subcarriers are generated, and the frequency offset interval of each subcarrier is set to the PN signal pulse width. These signals are input into the mixer after non-coherent accumulation.

[0161] 2. TCN-Seq2Seq model construction module: This module uses a 10-layer residual network to construct the TCN-Seq2Seq model. The input features include the circular convolution correlation output received from the local carrier compression module, the local DDC waveform output without frequency offset, and the original received waveform.

[0162] 3. Training module: offline training is performed using simulated waveform data for submarine communication, and a TCN-Seq2Seq capture model adapted to the submarine communication multipath channel environment is obtained.

[0163] 4. Post-processing module: This module post-processes the results of the TCN-Seq2Seq capture model, realizes time domain coarse synchronization under low signal-to-noise ratio conditions, obtains coarse estimation of Doppler frequency offset, and completes time domain and frequency domain coarse capture.

[0164] In both embodiments, Fine-Tune transfer learning can be performed based on the known prior information of the received waveform in the online test phase, to compensate for the differences between the simulation phase and the actual sea multipath environment and noise environment with small computational overhead.

[0165] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any modification, equivalent replacement and improvement within the technical range disclosed by the present application and within the spirit and principle of the present application should be covered within the protection scope of the present application.

Claims

1. A deep learning based time domain coarse acquisition method for underwater acoustic PN signals, characterized in that, Comprise the following steps: The first step is to convert the PN correlation gain amplitude frequency offset response into a plurality of SINC function module incoherent accumulation approximation flat based on the local carrier compression circular convolution time domain parallel correlation result output for providing the TCN-Seq2Seq model with large Doppler tolerance correlation waveform characteristics; The second step is to determine the input features of the TCN-Seq2Seq model, which include the circular convolution correlation output based on the local carrier compression, the local DDC waveform output without frequency offset, and the original received waveform; The TCN-Seq2Seq model is constructed to include 10 residual network blocks, global pooling layers, fully connected layers, and softmax layers, and finally outputs a classification sequence with the same length as the input time sequence, providing an analysis basis for subsequent time and frequency domain coarse synchronization; The third step is to complete the offline training based on the target sea area simulation waveform data, and obtain the TCN-Seq2Seq capture model suitable for the target sea area multipath channel environment; In the online test stage, first carry out Fine-Tune transfer learning based on the known prior information received waveform, and compensate for the differences between the simulation stage and the actual sea multipath environment and noise environment with low computational overhead by only training the last fully connected layer and classification layer of the TCN-Seq2Seq model, and freezing the front 10 residual network blocks and global pooling layers. The fourth step is to post-process the TCN-Seq2Seq capture model results to realize time domain coarse synchronization under low SNR conditions and obtain Doppler frequency offset coarse estimation to complete time and frequency domain coarse capture.

2. The deep learning-based underwater acoustic PN signal time-domain coarse acquisition method of claim 1, wherein, The first step of local carrier compression refers to generating C=2N+1 local sub-carriers in the process of down-conversion at the receiving end, entering into mixing through non-coherent accumulation, and spacing the frequency deviation of each sub-carrier by D f Taking 2 / T, T is the PN signal pulse width, and the subsequent is a circular convolution process, that is, frequency domain multiplication is equivalent to circular convolution, and the correlation of the zero-padded local baseband PN signal and the down-converted signal of the compressed carrier is completed, and finally the time domain parallel search synchronization is completed through judgment. 3.The deep learning based underwater acoustic PN signal time domain coarse acquisition method according to claim 2, wherein, The received signal is represented as: X r (t) = PN(t) - cos [2πf c (1 + β)t] + n Where: A denotes signal amplitude, PN code corresponds to chip sequence p = [p0p1p2...p L-1 ], L is PN code length, T c is chip width, f c is carrier frequency, β is Doppler factor, and the local compressed carrier is represented as: After mixing, we get: X m (t) = X r (t) · x l After low-pass filtering to remove high-frequency terms, we get: After entering the circular convolution, when the PN signal time domain is synchronized with the local baseband PN signal, the correlation result is: PN 2 (t) = 1, then: For the frequency offset Doppler factor β determined by the received signal, the correlation gain is the sum of 2N+1 SINC functions. 4.The deep learning based underwater acoustic PN signal time domain coarse acquisition method according to claim 1, wherein, The second step is to train the PN code capture synchronization from a sequence to sequence perspective. The PN code sequence position is random. The label of the time domain segment where the PN code exists is set to 1, and the label of the time domain segment where the PN code does not exist is set to 0. After normalizing the received signal, the label is added. The training set containing the PN code is referred to as the positive set, and part of the pure noise signal in the training set is set as the negative set. 5.The deep learning based underwater acoustic PN signal time domain coarse acquisition method according to claim 4, wherein, The ratio of the PN code time domain length to the pure noise time domain length in the positive set is 2:1, the sample time domain window length of the negative set is consistent with that of the positive set, and the ratio of the number of positive set samples to the number of negative set samples is 3:

1. In this way, the PN code time domain occupies 1:1 of the total training set, and the same proportion of sample data is used in the validation set and the test set. 6.The deep learning based underwater acoustic PN signal time domain coarse acquisition method according to claim 1, wherein, The underwater acoustic PN signal time domain coarse acquisition method further includes: TCN-Seq2Seq model sequence classification capability analysis, the target PN code passband signal time domain is expanded to 1.5T, the starting position of the PN code passband signal is randomly set and recorded, the time domain search step step span is 0.5T, and 1dB is taken as a step from-30dB to 0dB as a signal-to-noise ratio condition to add a white noise signal in the training set setting process; the training set is 3x10 -4 The Doppler factor is taken as a step and is from-2.1x10 -3 to 2.1x10 -3 A resampling simulation mobile environment is carried out; 300 groups of positive example training sets are extracted under each signal-to-noise ratio condition and Doppler condition, the training set labels are set according to the starting position of each group of PN codes and the PN code passband length, the corresponding position of the PN code is set as 1, and the remaining positions are set as 0; meanwhile, 100 groups of pure noise are generated under each signal-to-noise ratio and Doppler condition, 100 groups of negative example training sets are generated after feature extraction, and the negative example training set labels are all 0; there are 400 groups of training set input signals under each signal-to-noise ratio condition, the number of verification sets is 1 / 10 of the number of training sets, and the generation mode is consistent with that of the training set. Based on the above search and capture model, it has a time domain 1.5T step search and capture capability, that is, it has a time domain parallel search capability; According to the above model setting method, the PN signal with code length 2047 is used as the capture object to train the sequence to sequence classification model.

7. The deep learning-based underwater acoustic PN signal time-domain coarse acquisition method of claim 1, wherein, The water acoustic PN signal time domain coarse capture method further comprises: TCN-Seq2Seq model result post-processing and capture capability analysis, including The classification judgment logic value output by the TCN-Seq2Seq model is converted into a numerical format, and the average value of the classification results of each point and the T-1 points after the point in the range of [1:0.5*T] is taken as the CFAR model input observation value; The model output values of the sampling points outside the T-1 points after the target point in the detection window in STEP1 are replaced by the guardian cells in the traditional CFAR, and the detection is carried out after the false alarm rate is set; When there is a CFAR decision result, it is considered that there is a target PN signal in the window, and the central position of the CFAR decision output result is taken as the predicted time domain starting point. 8.A deep learning based underwater acoustic PN signal time domain coarse acquisition system, characterized in that, The system comprises: The local carrier compression module is used for converting the PN correlation gain amplitude frequency offset response into a plurality of SINC function module non-coherent accumulation approximate flat, and outputting the circular convolution time domain parallel correlation result based on the local carrier compression, so as to provide the TCN-Seq2Seq model with a large Doppler tolerance correlation waveform feature; The TCN-Seq2Seq model construction module is used for determining the input features of the TCN-Seq2Seq model, and the input features include the circular convolution correlation output based on the local carrier compression, the local DDC waveform output without frequency offset and the original received waveform; the module comprises 10 residual network blocks, a global pooling layer, a full connection layer and a softmax layer, and finally outputs a classification sequence with the same length as the input time sequence; The training module is used for offline training based on the target sea area simulation waveform data, so as to obtain a TCN-Seq2Seq capture model suitable for the target sea area multipath channel environment; The post-processing module is used for post-processing the TCN-Seq2Seq capture model result, realizing the time domain coarse synchronization under a low signal-to-noise ratio condition, obtaining a Doppler frequency offset coarse estimation, and completing the time domain and frequency domain coarse capture. 9.The deep learning based underwater acoustic PN signal time domain coarse acquisition system of claim 8, wherein, The training module performs Fine-Tune transfer learning based on the known prior information received waveform in the online test stage, compensates for the differences between the simulation stage and the actual sea multipath environment and noise environment by only training the last full connection layer and the classification layer of the TCN-Seq2Seq model, and freezing the front 10 residual network blocks and the global pooling layer.

10. The deep learning based underwater acoustic PN signal time domain coarse acquisition system of claim 8 or 9, wherein, The local carrier compression realized by the local carrier compression module refers to that a plurality of local subcarriers are generated in the receive end down-conversion process, enter the frequency offset interval for the PN signal pulse width, and then are subjected to the circular convolution process, that is, the frequency domain multiplication is equivalent to the circular convolution, the zero-padded local baseband PN signal is correlated with the signal after the compressed carrier down-conversion, and finally the time domain parallel search synchronization is completed through the decision.

Images