Underwater non-fixed node communication method based on deep learning and deep transfer learning

Abstract

The application belongs to the technical field of communication, and discloses an underwater non-fixed node M-element spread spectrum communication method based on TCN deep learning and Fine-tuning deep transfer learning, which is characterized in that the method specifically comprises the following steps: S1, TCN network model design; S2, TCN offline intensive training (pre-training); and S3, TCN transfer learning based on offshore measured data. In the underwater non-fixed node M-element spread spectrum communication in the complex environment of offshore Doppler and multipath, the application is more suitable for low signal-to-noise ratio and Doppler background sea conditions compared with traditional spread spectrum receiving technology and frequency compression technology. The underwater non-fixed node M-element spread spectrum communication method based on TCN deep learning and Fine-tuning deep transfer learning improves the communication effect of the traditional spread spectrum method. Simulation results show that, with a bit error rate of 0.01 as a reference, the signal-to-noise ratio of a 127 code TCN model can be reduced by 3 dB compared with that of a frequency compression receiver, and the signal-to-noise ratio of a 511 code TCN model can be reduced by 6 dB compared with that of a frequency compression receiver. Test results show that the TCN-TL model after transfer learning flexibly adapts to the noise and channel environment of the test sea area, and the decoding effect is obviously better than that of traditional technology.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of communication, and particularly relates to a kind of underwater non-fixed node M element spread spectrum communication method based on TCN deep learning and Fine-tuning deep migration learning. The method first applies the TCN deep learning model to the field of M element spread spectrum underwater acoustic communication, avoids the problems such as gradient explosion, gradient disappearance and large memory requirement caused by the long time sequence classification process of network structures such as LSTM and GRU in the existing recurrent neural network (RNN), and based on the characteristics that the convolutional neural network is not sensitive to the influence of Doppler on time sequence signal compression and expansion, a TCN network architecture suitable for underwater long spread spectrum code communication under complex Doppler background is explored through a large number of experiments. At the same time, aiming at the differences between the offline training process of simulation data and the actual sea communication process in terms of noise and channel environment, Fine-tuning deep migration learning training is carried out based on simulation data offline TCN model and part of training data on the sea, and the TCN-TL model obtained is verified by experiment to be obviously superior to the traditional direct sequence spread spectrum (DSSS) underwater acoustic communication system in terms of communication effect under the conditions of low signal-to-noise ratio, complex Doppler and time-varying multipath near the sea.

[0002] The above technology has not been published in the form of papers and the like, and fills the technical gap of the application of TCN deep learning model in the field of long code spread spectrum underwater acoustic communication. BACKGROUND

[0003] The robustness of underwater acoustic communication is significantly affected by the time-varying complex fluctuating dynamic environment of ocean currents, waves, internal waves and turbulence in the ocean. Due to the fact that the underwater sound speed is much lower than the speed of radio propagation in the air, under the same radial Doppler disturbance, the Doppler factor in the underwater acoustic channel is nearly 10 5 times that of the radio frequency radio channel, and the Doppler phenomenon has a significant impact on the communication effect in actual sea communication. In the near-shore underwater acoustic communication network based on the application scenario of underwater acoustic communication nodes of buoys, the communication nodes are relatively fixed by being anchored to the seabed through acoustic releasers and cables, but they will move relatively due to near-shore tidal currents, various mesoscale phenomena and random fluctuations. Improving the Doppler tolerance of the system is an important task to improve the robustness of communication in the process of building an underwater acoustic communication network.

[0004] M element DSSS can be used for underwater CDMA networking communication, and the bandwidth is expanded and the spectral energy density is reduced through pseudo-random code. Due to the characteristics of low underwater acoustic pollution (good biological friendliness) and low interception probability (good concealment security) caused by low transmission power, M element DSSS is applied to underwater Internet of Things networking communication. However, the spread spectrum communication of M element DSSS, especially long spread spectrum code, is extremely sensitive to the Doppler phenomenon, mainly in terms of carrier phase jump and time domain compression and expansion.

[0005] The prior art often adopts the following technical scheme:

[0006] (1)Traditional spread spectrum receiving technology. After band-pass filtering and time-domain synchronization, the receiving end realizes carrier phase synchronization through phase-locked loop technology, mixes with the local carrier oscillator, and then low-pass filters to complete down-conversion and obtain the baseband signal. Based on the spread spectrum code index library, the baseband signal is matched and correlated in turn, and the maximum value method is used to obtain the spread spectrum code index, and then decoding is completed.

[0007] (2)Frequency compression receiving technology. After band-pass filtering and time-domain synchronization, the receiving end develops multiple down-conversion frequencies through local grid frequencies, each of which is low-pass filtered to complete carrier removal, and the non-coherent superposition of each signal is correlated with M local spread spectrum codes, and the maximum value of the spread spectrum gain value is selected to complete symbol classification judgment, and then decoding is completed.

[0008] The traditional spread spectrum receiving technology realizes carrier synchronization through phase-locked loop technology, and dynamically tracks the changes of carrier frequency and phase after completing acquisition and locking. However, its performance decreases significantly at low signal-to-noise ratio; the frequency compression receiving technology loses signal gain in the non-coherent superposition process, and has limited effect on improving the system's ability to resist Doppler, and increases the computational overhead and reduces the service life of the underwater node. SUMMARY

[0009] In view of the problems existing in the prior art, the present application provides a kind of underwater non-fixed node M element spread spectrum communication method based on TCN deep learning and Fine-tuning deep migration learning.

[0010] The present application is realized in this way, a kind of underwater non-fixed node M element spread spectrum communication method based on TCN deep learning and Fine-tuning deep migration learning, this method specifically includes:

[0011] S1: TCN network model design;

[0012] S2: TCN offline intensive training (pre-training);

[0013] S3: TCN migration learning based on sea measured data.

[0014] Further, the S1 specifically includes:

[0015] According to the characteristics of M element underwater acoustic communication spread spectrum receiving waveform signal, a TCN deep learning network model based on residual network is designed, an exponential increase of receptive field is obtained based on dilated convolution and deep residual network along with the increase of network depth, in the actual TCN network, the base of dilated convolution is set to 2, and the receptive field increases exponentially with the increase of layer number; at the same time, in order to further expand the receptive field, the convolution kernel size is set to 5;After the TCN residual block, a global pooling layer, a fully connected layer, a Softmax layer and a Classification layer are set to classify.

[0016] Further, the S2 specifically comprises:

[0017] The TCN model is trained offline using the BELLHOP simulation signal line, and during the training process, the data enhancement preprocessing process for the Doppler and multipath is added to the simulation training set to strengthen the preprocessing process for the Doppler and multipath under the actual complex power environment and strong time-varying channel conditions.

[0018] Further, the data enhancement preprocessing process for the Doppler and multipath is specifically as follows:

[0019] The training data is strengthened by inputting the data of -1 to 1 m / s Doppler background and the 0.2 m / s grid channel background, and white noise with a signal-to-noise ratio of -10 dB is added; at the same time, the spatial grid training node is set, the simulation signal strengthening training based on the simulation communication channel between the grid points is carried out, and the spatial generalization ability of the TCN model is improved; the green random test node in the spatial grid training node schematic effect is used to carry out the simulation data online test in random position and random Doppler background.

[0020] Further, the S3 specifically comprises:

[0021] The TCN model trained based on the simulation data directly applied to the actual observation data of the sea test produces a bit error rate platform effect;

[0022] For the noise and channel mismatch, Fine-tuning deep transfer learning is adopted, the TCN model trained based on the simulation data is used as the pre-training model, the convolution residual network before the global pooling layer is frozen, part of the test data is input into the TCN network as the agreed communication training data, and the transfer learning is carried out on the full connection layer and the Softmax layer at the back end of the network;

[0023] The communication effect of the TCN-TL model under the low signal-to-noise ratio condition is evaluated, the environmental noise signal collected in the test is adjusted in amplitude according to the signal-to-noise ratio energy condition, and is superimposed on the test received waveform signal to obtain the examination waveform signal under the low signal-to-noise ratio condition, the remaining data is input into the model (TCN-TL) after the transfer learning, and the decoding bit error rate is compared with the traditional spread spectrum receiver, the frequency compression receiver and the TCN receiver based on the simulation data training.

[0024] Further, in the S3, the bit error rate platform effect produced when the TCN model trained based on the simulation data is directly applied to the actual observation data of the sea test specifically comprises:

[0025] Based on the 511 code M communication receiving waveform data obtained in July 2022 in the relative radial velocity of 0.32 m / s of the transceiver node in Jiaozhou Bay, the decoding error rates of the traditional spread spectrum receiver, the frequency compression receiver and the TCN receiver based on the simulation data training are 0.0788, 0.0363 and 0.0863 respectively; the white noise used in the offline training process is relatively approximate to simulate the offshore environmental noise, and the actual offshore ocean environmental noise is complex, including a large amount of completely random white noise, as well as transient noise, colored noise and other signals caused by biology and human beings; in terms of channel simulation, due to the difficulty in accurately reflecting the complex topography, geology and rough sea surface and other environmental conditions, and the existence of operation errors of the underwater acoustic channel model, the offline simulation channel condition also has errors.

[0026] Further, the Fine-tuning deep transfer learning specifically includes:

[0027] Based on 800 groups of measured waveform data in Jiaozhou Bay test, 200 groups of which are used for Fine-tuning deep transfer learning, the communication effect of TCN-TL model under low signal-to-noise ratio condition is evaluated, the environmental noise signal collected in the test is adjusted in amplitude according to the signal-to-noise ratio energy condition, and is superimposed to the test receiving waveform signal to obtain the examination waveform signal under low signal-to-noise ratio condition; then the remaining data is introduced into the model (TCN-TL) after transfer learning, and the decoding error rates of the traditional spread spectrum receiver, the frequency compression receiver and the TCN receiver based on the simulation data training are compared.

[0028] In combination with the above technical solutions and the technical problems solved, the technical solutions to be protected by the present application have the following advantages and positive effects:

[0029] First, in view of the technical problems existing in the above prior art and the difficulty in solving the problems, the present application closely combines the technical solutions to be protected and the results and data in the research and development process, and analyzes in detail and deeply how the technical solutions solve the technical problems and bring some creative technical effects after solving the problems. The specific description is as follows:

[0030] This invention proposes an underwater non-fixed node M-ary spread spectrum communication method based on the TCN deep learning model. This method designs a TCN-structured deep learning network based on the residual network principle, which, after training, acts as a receiver processor to identify and classify underwater acoustic spread spectrum time-series waveform signals, thereby completing decoding without consuming significant computational resources for channel estimation, channel equalization, and despreading correlation calculations. The TCN model training process consists of two stages: offline simulation reinforcement training and actual sea transfer learning. In the offline simulation reinforcement training stage, offline reinforcement training is conducted using a spatial network grid BELLHOP simulated channel and a Doppler spread background to achieve high generalization capability of TCN in both spatial and Doppler environments. In the actual sea transfer learning stage, to address the mismatch between multipath and noise environments, a Fine-tuning deep transfer learning training method is proposed. This method freezes the convolutional residual network before the global pooling layer, assumes some experimental data as pre-communication training data, and inputs it into the Fine-tuning TCN network to perform transfer learning on the fully connected layers and Softmax layers at the back end of the network. Simulations and experiments show that this method significantly improves the performance of M-element spread spectrum underwater acoustic communication in complex Doppler multipath environments near the coast.

[0031] This invention, compared to traditional spread spectrum receiving and frequency compression techniques, is more suitable for underwater non-fixed node M-ary underwater acoustic spread spectrum communication in complex nearshore Doppler and multipath environments, particularly in low signal-to-noise ratio (SNR) and Doppler background sea conditions. The underwater non-fixed node M-ary spread spectrum communication method proposed in this invention, based on TCN deep learning and fine-tuning deep transfer learning, improves the communication performance of traditional spread spectrum methods. Simulation results show that, with a bit error rate of 0.01%, the 127-code TCN model can reduce the SNR by 3 dB compared to the frequency compression receiver, and the 511-code TCN model can reduce the SNR by 6 dB compared to the frequency compression receiver. Sea trials demonstrate that the transfer-learned TCN-TL model flexibly adapts to the noise and channel environment of the test sea area, and its decoding performance is significantly better than traditional techniques.

[0032] Second, considering the technical solution as a whole or from a product perspective, the technical effects and advantages of the technical solution to be protected by this invention are specifically described as follows:

[0033] The present application aims at the problems of phase jump and sharp decline of spread spectrum gain of traditional spread spectrum communication in underwater mobile communication, uses the classification function of a deep learning time convolution network (TCN) model as a receiver of an M-element spread spectrum communication system to complete spread spectrum code classification and then realize decoding. In the training process, the neural network model based on the TCN structure is input with the time domain waveform of each chaotic sequence to learn the features carried by each sequence; through the spatial network grid BELLHOP simulation channel and the Doppler spread background, offline reinforcement training is carried out to realize the high generalization ability of TCN in space and Doppler; in the online application process in the actual marine environment, for the mismatching conditions of multipath environment and noise environment, a Fine-tuning deep transfer learning training method is proposed, the convolution residual network before the global pooling layer is frozen, part of the test data is set as the agreed known pre-training data, and the Fine-tuning TCN network is input to realize transfer learning of the full connection layer and the Softmax layer at the back end of the network, thereby improving the robustness of the M-element spread spectrum communication under the complex Doppler and time-varying multipath environment.

[0034] As a product application technical solution, the communication receiver based on the TCN deep learning model can improve the communication effect in the low signal-to-noise ratio complex marine environment; meanwhile, the channel estimation, channel equalization and despread correlation calculation links in the traditional receiving system are omitted, the computing cost of the underwater acoustic communication node is reduced, and the service life of the underwater communication node is prolonged; compared with the application scheme of the internationally leading GNN model in the field of underwater acoustic spread spectrum communication in recent years, the application of the TCN deep learning model in the field of M-element spread spectrum communication breaks through the application limitation of deep learning in the field of long spread spectrum code, and solves the problems of gradient explosion and gradient disappearance in the network learning process.

[0035] Thirdly, the creativity of the present application as the claim is also embodied in the following important aspects:

[0036] The technical solution of the present application fills the domestic and foreign technical blank:

[0037] The method applies the TCN deep learning model to the field of M-element spread spectrum underwater acoustic communication for the first time, further breaks through the application limitation of deep learning in the field of long spread spectrum code on the basis of the application scheme of the internationally leading GNN model in the field of underwater acoustic spread spectrum communication in recent years, and solves the problems of gradient explosion and gradient disappearance in the network learning process. The TCN convolutional neural network is not sensitive to the compression and expansion of time series, and the robustness problem of underwater acoustic spread spectrum communication under complex Doppler background is better solved. The deep learning in the field of underwater acoustic spread spectrum communication is applied to practical sea test verification for the first time, the sea Fine-tuning transfer learning with small amount of calculation is realized, the sea communication effect verification is realized, and the communication effect is obviously better than that of the existing underwater acoustic spread spectrum communication system. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flow chart of an underwater non-fixed node M-spread spectrum communication method based on TCN deep learning and Fine-tuning deep transfer learning provided by an embodiment of the present application.

[0039] Figure 2 is a TCN network dilated convolution schematic diagram provided by an embodiment of the present application.

[0040] Figure 3 is a TCN deep residual network structure provided by an embodiment of the present application.

[0041] Figure 4 is a spatial grid training node schematic diagram (environmental conditions consistent with Jiaozhou Bay test) provided by an embodiment of the present application.

[0042] Figure 5 is a bit error rate (BER) performance comparison chart of a TCN receiver, a traditional SS receiver (Con-S) and a frequency compression receiver (Fre-C) based on simulation test data provided by an embodiment of the present application.

[0043] Figure 6 is a Fine-tuning deep transfer learning process diagram provided by an embodiment of the present application.

[0044] Figure 7 is a bit error rate (BER) performance comparison chart of a TCN-TL receiver, a TCN receiver, a traditional SS receiver (Con-S) and a frequency compression receiver (Fre-C) based on sea test data provided by an embodiment of the present application. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below 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.

[0046] In order for those skilled in the art to fully understand how the present application is specifically implemented, this part is an explanatory description of the embodiments of the technical scheme of the claims.

[0047] As shown in Figure 1 , the underwater non-fixed node M-spread spectrum communication method based on Temporal Convolutional Network (TCN) deep learning and Fine-tuning deep transfer learning provided by the present application specifically includes the following steps:

[0048] S1: TCN network model design;

[0049] S2: TCN offline intensive training (pre-training);

[0050] S3: TCN based on offshore measured data migration learning.

[0051] In order to solve the problem of M-element spread spectrum robust communication in the near-shore underwater complex Doppler and time-varying multipath environment, the application provides an underwater non-fixed node M-element spread spectrum communication method based on a TCN deep learning model. The method is based on the principle of TCN deep learning model and designs a TCN structure deep learning network based on residual network. After training, the TCN structure deep learning network is used as a receiver processor to identify and classify underwater acoustic spread spectrum time sequence waveform signals, and then decoding is completed without consuming a large amount of computing resources to carry out channel estimation, channel equalization, and de-spreading correlation calculation. The TCN model training process includes offline simulation reinforcement training and actual offshore migration training. In the offline simulation reinforcement training stage, the spatial network grid BELLHOP simulation channel and Doppler spread background are used for offline reinforcement training, so that the TCN has high generalization ability in space and Doppler. In the actual offshore migration learning stage, the Fine-tuning deep migration learning training method is used to freeze the convolutional residual network before the global pooling layer, set part of the test data as the agreed known pre-training data, and input the Fine-tuning TCN network to migrate the learning of the full connection layer and the Softmax layer at the back end of the network. Simulations and tests show that the method significantly improves the effect of M-element spread spectrum underwater acoustic communication in the near-shore complex Doppler and multipath environment. The method steps are described as follows:

[0052] 1. TCN network model design.

[0053] According to the characteristics of the M-element underwater acoustic communication spread spectrum received waveform signal, a TCN deep learning network model based on residual network is designed. The TCN dilated convolution schematic diagram is shown in Figure 2 , the TCN deep residual network structure is shown in Figure 3 , and the TCN network design parameters are shown in Table 1. In the actual TCN network in this paper, the dilated convolution base is set to 2, and the receptive field increases exponentially with the number of layers; at the same time, in order to further expand the receptive field, the convolution kernel size is set to 5. After the TCN residual block, the global pooling layer, the full connection layer, the Softmax layer and the Classification layer are set for classification. It should be pointed out that through repeated tests, it is found that the original data received and the data processed by the carrier frequency fc down frequency (without considering the Doppler factor) are used as two feature inputs of the TCN, which obtains better learning effect than using only the original data as the feature input, which may be due to the system obtaining the carrier frequency information of the transmitted signal, thereby improving the feature extraction efficiency.

[0054] Table 1 sets the design parameters for the M-element underwater acoustic spread spectrum communication TCN model

[0055]

[0056]

[0057] 2. TCN offline intensive training (pre-training)

[0058] The TCN model is trained offline using the BELLHOP simulation signal. During the training process, in order to deal with the radial Doppler and strong time-varying nature of the channel conditions in the actual complex dynamic environment, we add data enhancement preprocessing for Doppler and multipath in the offline training set: the training data is strengthened by inputting -1 to 1 m / s Doppler background data, 0.2 m / s grid channel background, and adding -10 dB signal-to-noise ratio white noise; At the same time, the spatial grid training nodes are set up, and the simulation signal is strengthened by carrying out simulation communication channel based on grid points, which improves the spatial generalization ability of the TCN model. The spatial grid training node is shown in Figure 4 . Based on the green random test nodes shown in Figure 4 , the simulation data online test is carried out in random position and random Doppler background, and two types of chaotic spread spectrum codes with lengths of 127 and 511 are set respectively. The test results obtained by Monte Carlo experiment are shown in Figure 5 . It can be seen that the TCN model has obvious improvement in communication effect under the conditions of random position and Doppler compared with traditional spread spectrum and frequency compression spread spectrum. Among them, the 511 code TCN model has more waveform features due to its longer code length, and has certain advantages over the 127 code TCN model. Taking the bit error rate of 0.01 as a reference, the signal-to-noise ratio of the 127 code TCN model can be reduced by 3 dB compared with the frequency compression receiver, and the signal-to-noise ratio of the 511 code TCN model can be reduced by 6 dB compared with the frequency compression receiver. The signal-to-noise ratio of the 511 code TCN model can be reduced by nearly 4 dB compared with the 127 code TCN model.

[0059] 3. TCN based on migration learning of sea measured data

[0060] We based on the 511 code M communication received waveform data obtained in July 2022 in Jiaozhou Bay in the Doppler background of the relative radial velocity of 0.32 m / s, and obtained the decoding error rate of the traditional spread spectrum receiver, the frequency compression receiver and the TCN receiver based on the simulation data training as 0.0788, 0.0363 and 0.0863 respectively. It can be seen that the TCN receiver based on the simulation data training fails to achieve the purpose of improving the effect of marine communication. It should be pointed out that the offline training process uses white noise to simulate the near-sea environmental noise, which is relatively approximate. The actual near-sea environmental noise is complex, and in addition to a large amount of completely random white noise, it also contains transient noise, colored noise and other signals caused by biology and human beings; in terms of channel simulation, due to the difficulty in accurately reflecting the complex topography, geology and rough sea surface and other environmental conditions, and the existence of operation errors of the underwater acoustic channel model, the offline simulation channel condition also has errors. Influenced by the above two points, when the TCN model trained based on the simulation data is directly applied to the actual observation data in the sea test, the error rate platform effect is generated.

[0061] In view of the mismatch of noise and channel, Fine-tuning deep transfer learning is proposed, the TCN model trained based on simulation data is used as the pre-training model, the convolution residual network before the global pooling layer is frozen, part of the test data is used as the agreed known training data before communication, and the TCN network is input to the transfer learning of the full connection layer and the Softmax layer at the back end of the network. The Fine-tuning deep transfer learning process is as shown in Figure 6 . Here, based on 800 groups of measured waveform data in Jiaozhou Bay test, we use 200 groups of them to carry out Fine-tuning deep transfer learning. In order to evaluate the communication effect of TCN-TL model under low signal-to-noise ratio condition, we adjust the amplitude according to the signal-to-noise ratio energy condition, and superimpose it on the test received waveform signal to obtain the test waveform signal under low signal-to-noise ratio condition. Then the remaining data is input into the model after transfer learning (TCN-TL), and the decoding error rate of the traditional spread spectrum receiver, the frequency compression receiver and the TCN receiver based on the simulation data training is compared Figure 7 . It can be seen that the TCN-TL model after transfer learning adapts to the noise and multipath environment of the test site, and achieves similar Figure 5 effect as the simulation data test, which significantly improves the communication ability of the system under complex near-sea dynamic environment, noise environment and low signal-to-noise ratio condition. The results of the sea test confirm the effectiveness of the method.

[0062] In order to prove the creativity and technical value of the technical scheme of the application, this part is the application embodiment of the technical scheme of the claim on the specific product or related technology.

[0063] The underwater non-fixed node M-ary spread spectrum communication method based on TCN deep learning and Fine-tuning deep migration learning provided by the application embodiment is applied to a computer device, the computer device comprising a memory and a processor, the memory storing a computer program, and the computer program is executed by the processor to make the processor execute the steps of the underwater non-fixed node M-ary spread spectrum communication method based on TCN deep learning and Fine-tuning deep migration learning.

[0064] The underwater non-fixed node M-ary spread spectrum communication method based on TCN deep learning and Fine-tuning deep migration learning provided by the application embodiment is applied to an information data processing terminal, and the information data processing terminal is used to realize the steps of the underwater non-fixed node M-ary spread spectrum communication method based on TCN deep learning and Fine-tuning deep migration learning.

[0065] The underwater non-fixed node M-ary spread spectrum communication method based on TCN deep learning and Fine-tuning deep migration learning provided by the application embodiment is applied to a computer readable storage medium, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the underwater non-fixed node M-ary spread spectrum communication method based on TCN deep learning and Fine-tuning deep migration learning.

[0066] It should be noted that the embodiments of the present application can be realized by hardware, software or a combination of software and hardware. The hardware part can be realized by special logic; the software part can be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or a specially designed hardware. Those skilled in the art can understand that the above-mentioned devices and methods can be realized by computer executable instructions and / or included in processor control code, such as provided on a carrier medium, such as a magnetic disk, CD or DVD-ROM, a programmable memory, such as a read-only memory (firmware), or a data carrier, such as an optical or electronic signal carrier. The devices of the present application and their modules can be realized by hardware circuits, such as very large scale integrated circuits or gate arrays, semiconductors, such as logic chips, transistors, etc., or programmable hardware devices, such as field programmable gate arrays, programmable logic devices, etc., can also be realized by software executed by various types of processors, and can also be realized by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0067] The application embodiment has achieved some positive effects in research and development or use, and indeed has great advantages compared with the prior art. The following content is described in combination with the data and graphs of the test process.

[0068] A water acoustic direct sequence spread spectrum communication experiment was carried out in Jiaozhou Bay on July 15, 2022. The tide type in Jiaozhou Bay is mainly semi-diurnal tide, and the day is the spring tide, so the tidal current is relatively strong. The signal transmitting ship was in the state of anchoring during the communication signal transmission process, and the initial release depth of the transmitting transducer was 15 meters. The signal receiving ship was located at the entrance of Jiaozhou Bay (the relative tidal current was relatively large), and the initial release depth of the receiving hydrophone was 15 meters. During the experiment, the ship was in a free moving drift state without anchoring, so there was a relative movement between the transmitting transducer and the receiving hydrophone, i.e. Doppler background, and the average radial relative velocity was 0.32 m / s. The radial distance between the signal transmitting ship and the receiving ship was 4000 meters, and the radial water depth changed relatively slowly, with a water depth of 28-34 meters. A total of 800 groups of spread spectrum code elements were sent.

[0069] The received signals in the sea trial were decoded by four methods: traditional spread spectrum receiver, frequency compression receiver, TCN model (trained based on simulation data), and TCN-TL (trained based on 200 groups of received waveform signals through Fine-tuning transfer learning). The bit error rate results are shown in FIG. 2, which shows that the TCN-TL receiver through transfer learning significantly improves the sea communication capability. The offline TCN model without transfer learning has a large difference between the ideal white noise and the simulation multipath channel in the offline training stage and the actual environment at sea, resulting in a bit error platform effect of the TCN model. Figure 7

[0070] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any modification, equivalent replacement and improvement within the technical range disclosed in the present application, which is within the spirit and principles of the present application, should be covered within the protection scope of the present application.​

Claims

1. An underwater non-fixed node M-ary spread spectrum communication method based on TCN deep learning and Fine-tuning deep transfer learning, characterized in that, Comprise: S1: TCN network model design; S2: TCN offline intensive training; S3: TCN migration learning based on sea measured data; The S1 specifically comprises: For the characteristics of M-element underwater acoustic communication spread spectrum receiving waveform signal, a TCN deep learning network model based on residual network is designed, an exponential increase in receptive field is obtained with the increase of network depth based on dilated convolution and deep residual network, in the actual TCN network, the base of dilated convolution is set to 2, the receptive field increases exponentially with the increase of layer number; At the same time, in order to further expand the receptive field, the convolution kernel size is set to 5; Global pooling layer, full connection layer, Softmax layer and Classification layer are set after TCN residual block for classification; The S2 specifically comprises: TCN model is trained by BELLHOP simulation signal, during the training process, for the radial Doppler in actual complex dynamic environment and strong time-varying nature of channel conditions, data enhancement preprocessing process for Doppler and multipath is added in the simulation training set in the offline training stage.

2. The underwater non-fixed node M-ary spread spectrum communication method based on TCN deep learning and Fine-tuning deep transfer learning according to claim 1, wherein, The data enhancement preprocessing process for Doppler and multipath is specifically as follows: The training data is strengthened by inputting-1 to 1m / s Doppler background data, 0.2m / s interval gridding channel background, and adding-10dB signal-to-noise ratio condition white noise; At the same time, the spatial grid training node is set, the spatial generalization ability of TCN model is improved by carrying out simulation signal strengthening training based on the simulation communication channel between grid points, and the green random test node in the spatial grid training node schematic effect is carried out. Random position, random Doppler background simulation data online test. 3.The underwater non-fixed node M-ary spread spectrum communication method based on TCN deep learning and Fine-tuning deep transfer learning of claim 1, wherein, The S3 specifically comprises: The TCN model trained based on simulation data directly applied to the actual observation data of sea test produces bit error rate platform effect; For the case of noise and channel mismatch, Fine-tuning deep migration learning is adopted, the TCN model trained based on simulation data is used as the pre-trained model, the convolution residual network before global pooling layer is frozen, part of the test data is set as the agreed known communication training data, and the TCN network is input to the network backend full connection layer and Softmax layer for migration learning; The communication effect of TCN-TL model under low signal-to-noise ratio condition is evaluated, the environmental noise signal collected in the test is adjusted in amplitude according to the signal-to-noise ratio energy condition, and is superimposed to the test receiving waveform signal to obtain the test waveform signal under low signal-to-noise ratio condition. The remaining data is imported into the model after migration learning, and the decoding bit error rate of traditional spread spectrum receiver, frequency compression receiver and TCN receiver based on simulation data training is compared.

4. The underwater non-fixed node M-ary spread spectrum communication method based on TCN deep learning and Fine-tuning deep transfer learning according to claim 3, wherein, In the S3, the TCN model trained based on simulation data directly applied to the actual observation data of sea test produces bit error rate platform effect, specifically comprising: Based on the obtained 511 code M-element communication receiving waveform data, the decoding bit error rates of traditional spread spectrum receiver, frequency compression receiver and TCN receiver based on simulation data training are 0.0788, 0.0363 and 0.0863 respectively.

5. The underwater non-fixed node M-ary spread spectrum communication method based on TCN deep learning and Fine-tuning deep transfer learning according to claim 3, wherein, The Fine-tuning deep migration learning specifically comprises: Based on the measured waveform data, Fine-tuning deep transfer learning is carried out, and the communication effect of the TCN-TL model under low signal-to-noise ratio conditions is evaluated. The environmental noise signals collected in the test are adjusted in amplitude according to the signal-to-noise ratio energy condition, and are superimposed on the test received waveform signals to obtain the test waveform signals under low signal-to-noise ratio conditions. Then, the remaining data is imported into the TCN-TL model after transfer learning, and the bit error rate is compared with that of the traditional spread spectrum receiver, the frequency compression receiver and the TCN receiver trained based on simulation data.

6. A computer device, comprising: The computer device comprises a memory and a processor, and the memory stores a computer program, and the computer program is executed by the processor to enable the processor to execute the steps of the underwater non-fixed node M-element spread spectrum communication method based on TCN deep learning and Fine-tuning deep transfer learning according to any one of claims 1-5. 7.A computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to enable the processor to execute the steps of the underwater non-fixed node M-element spread spectrum communication method based on TCN deep learning and Fine-tuning deep transfer learning according to any one of claims 1-5.

8. An information data processing terminal, characterized by The information data processing terminal is used to realize the steps of the underwater non-fixed node M-element spread spectrum communication method based on TCN deep learning and Fine-tuning deep transfer learning according to any one of claims 1-5.

Images