Space-time structured sparse light coding deep water heterogeneous network data synchronization method and system

Abstract

The application provides a kind of spatio-temporal structure sparse light coding deep water heterogeneous network data synchronization method and system, including based on the overcomplete dictionary trained, the time sequence clock deviation data of each underwater node in the sending terminal equipment of deep water heterogeneous network is sparsely decomposed, and the original sparse vector is obtained;After the dynamic compressed observation data corresponding to the original sparse vector is converted into optical signal, it is transmitted to the receiving terminal equipment of deep water heterogeneous network by umbilical cable by sending terminal equipment;After the optical signal is demodulated into observation vector by receiving terminal equipment, it is input into the trained deep unfolding network model, and the reconstructed sparse vector is output by deep unfolding network model;The reconstructed sparse vector is recovered into clock deviation signal by overcomplete dictionary, and clock synchronization calibration instruction is generated, to realize the data synchronization of corresponding underwater node.The application can solve the problems of high bandwidth occupation, low signal ratio and poor reconstruction timeliness in underwater network time synchronization technology.

Description

Technical Field

[0001] This invention relates to the fields of deepwater oil and gas engineering and underwater precision measurement technology, and in particular to a spatiotemporally structured sparse optically coded deepwater heterogeneous network data synchronization method and system. Background Technology

[0002] As deep-sea oil and gas development moves towards digitalization and intelligence, Subsea Production Systems (SPS) integrate a vast number of environmental monitoring, structural health monitoring, and fluid control modules. These modules are distributed across a wide seabed area, forming a huge underwater heterogeneous network, whose collaborative operation and data fusion highly depend on nanosecond-level high-precision time synchronization.

[0003] However, existing underwater network time synchronization technologies, such as fiber optic Ethernet synchronization methods based on PTP (Precision Time Protocol) / NTP (Network Time Protocol), face the following serious challenges:

[0004] (1) Traditional PTP / NTP protocols broadcast the full time frame and have high redundancy. In large-scale underwater sensor networks, they are prone to occupying bandwidth, causing congestion, and affecting high-bandwidth services.

[0005] (2) The transmission loss and interference of deep-water umbilical cables cause the signal-to-noise ratio of traditional IM-DD (Intensity Modulation-Direct Detection) signals to decrease, resulting in synchronization clock jitter.

[0006] (3) The measurement matrix of the traditional compressed sensing (CS) scheme is not adapted to the underwater channel, and the traditional reconstruction algorithm (such as OMP, Orthogonal Matching Pursuit) is time-consuming to iterate, which is difficult to meet the real-time requirements of industrial control.

[0007] There is currently no effective solution to the problems of high bandwidth consumption, low signal-to-weight ratio, and poor reconfiguration timeliness in underwater network time synchronization technology. Summary of the Invention

[0008] The spatiotemporal structured sparse optical coding deep-water heterogeneous network data synchronization method and system provided by the present invention at least solves the problems of high bandwidth consumption, low signal-to-weight ratio and poor reconstruction timeliness in underwater network time synchronization technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] The first aspect of this invention provides a spatiotemporally structured sparse optically coded data synchronization method for deep-water heterogeneous networks, comprising the following steps: based on a trained overcomplete dictionary, sparsely decomposing the timing clock deviation data of each underwater node using a transmitting device of the deep-water heterogeneous network to obtain an original sparse vector; converting the dynamically compressed observation data corresponding to the original sparse vector into an optical signal, and then transmitting the optical signal to a receiving device of the deep-water heterogeneous network via an umbilical cable using the transmitting device; wherein the dynamically compressed observation data is a dynamic chaotic observation matrix generated from the timing clock deviation data. The original sparse vector is obtained by compressed sampling; after the optical signal is demodulated into an observation vector by the receiving device, it is input into a trained deep unfolded network model, and the deep unfolded network model outputs a reconstructed sparse vector; wherein, the deep unfolded network model is based on a fixed-layer neural network architecture, and is obtained by end-to-end training of each layer of the network using training pairs consisting of simulated sparse vectors and noisy chaotic observation vectors; the reconstructed sparse vector is restored to a clock deviation signal by the overcomplete dictionary, and a clock synchronization calibration command is generated to realize the data synchronization of the corresponding underwater node.

[0011] Preferably, before performing sparse decomposition on the timing clock skew data of each underwater node through the transmitting device of the deep-water heterogeneous network, the method includes the following steps: based on a randomly generated initial dictionary, calculating a sparse coefficient matrix that minimizes the reconstruction error between the reconstructed sample matrix and the training sample set, and satisfies sparsity constraints; wherein, the training sample set includes timing clock skew data of multiple underwater nodes within multiple time periods of the deep-water heterogeneous network; the reconstructed sample matrix is ​​obtained by multiplying the initial dictionary and the sparse coefficient matrix; based on the non-zero coefficients of the sparse coefficient matrix, determining the sample index corresponding to the initial dictionary, and constructing a restricted residual matrix corresponding to the sample index; based on the new atoms and new coefficients obtained by performing singular value decomposition on the restricted residual matrix, iteratively updating the initial dictionary and the sparse coefficient matrix until the change in the reconstruction error is less than a preset threshold or the maximum number of iterations is reached, thereby obtaining a trained overcomplete dictionary.

[0012] Preferably, constructing the restricted residual matrix corresponding to the sample index includes the following steps: constructing the residual matrix corresponding to each column atom based on the initial dictionary; wherein, the residual matrix is ​​constructed based on the difference between the training sample set and the corresponding reconstruction matrix; the reconstruction matrix is ​​the sum of multiplying the atoms in other columns (excluding the current column atom) with the corresponding coefficient row vectors in the sparse coefficient matrix; and selecting the corresponding restricted residual matrix from each residual matrix based on the sample index.

[0013] Preferably, the initial dictionary and sparse coefficient matrix are iteratively updated based on the new atoms and new coefficients obtained by performing singular value decomposition on the restricted residual matrix, including the following steps: performing singular value decomposition on the restricted residual matrix to obtain the corresponding left singular vector matrix, diagonal matrix, and right singular matrix; generating corresponding new atoms and new coefficients based on the left singular vector matrix, the diagonal matrix, and the right singular matrix; wherein, the new atom is the first column of the left singular vector matrix; and the new coefficient is the product of the first element of the diagonal matrix and the first column of the right singular matrix.

[0014] Preferably, converting the dynamically compressed observation data corresponding to the original sparse vector into an optical signal includes the following steps: generating a chaotic initial value based on the current timestamp of the time-series clock deviation data through a unidirectional irreversible numerical mapping operation; wherein the initial value is in the range of 0 to 1; inputting the initial value into the chaotic mapping equation for iterative calculation to generate a dynamic chaotic observation matrix that satisfies a preset length; performing compressed sampling on the original sparse vector based on the dynamic chaotic observation matrix to obtain dynamically compressed observation data; and converting the dynamically compressed observation data into an optical signal based on the mapping relationship between the electrical signal and the phase difference of adjacent optical pulses.

[0015] Preferably, the initial value is input into the chaotic mapping equation for iterative calculation to generate a dynamic chaotic observation matrix that meets a preset length, including the following steps: inputting the initial value into the chaotic mapping equation for iterative calculation, discarding the transient values ​​generated by the previous set number of iterations, and generating a pseudo-random sequence of a preset length; rearranging the purified pseudo-random sequence into an initial matrix of a preset dimension according to row priority or column priority rules; normalizing the column vectors of the initial matrix, and then optimizing the orthogonality of the column vectors through vector projection decomposition to obtain the final dynamic chaotic observation matrix.

[0016] Preferably, before inputting the observation vector into the trained deep unfolded network model, the method further includes the following steps: constructing a deep unfolded network model based on a neural network architecture with a fixed number of layers; wherein the operation rules of the deep unfolded network model include: fusing the received observation information with the reconstruction results of the previous layer based on a learnable weight matrix, and then obtaining the reconstruction results of each layer of the network through soft thresholding; and performing end-to-end training on each weight matrix and the threshold parameters corresponding to the soft thresholding calculation based on training pairs composed of randomly generated simulated sparse vectors that satisfy sparsity constraints and noisy chaotic observation vectors, combined with a loss function, to obtain a trained deep unfolded network model.

[0017] Preferably, the expression for the operation rule of the k-th layer of the deep unfolded network is: ;in, This represents the output of the (k+1)th layer of the network; This represents the output of the k-th layer network. This represents the observation information received by the depth-unfolded network; This represents the first weight matrix corresponding to the observation information of the k-th layer network; This represents the second weight matrix corresponding to the output of the k-th layer of the network; This represents the soft-threshold activation function value of the k-th layer network. , This represents the input variable for the soft-threshold activation function; Represents a symbolic function; This represents the threshold parameter of the k-th layer network; This represents the maximum value function.

[0018] Preferably, the process of restoring the reconstructed sparse vector to a clock bias signal using the overcomplete dictionary and generating a clock synchronization calibration command to calibrate the clock of the corresponding underwater node includes the following steps: restoring the reconstructed sparse vector to a clock bias signal using the overcomplete dictionary; comparing the clock bias signal with a bias threshold, and generating a clock synchronization calibration command when the absolute value of the clock bias signal is greater than the bias threshold; wherein the clock synchronization calibration command includes: an identifier of the target underwater node and a calibration value; and sending the clock synchronization calibration command via the receiving device to the corresponding target underwater node through the underwater local area network to achieve data synchronization of the corresponding underwater node.

[0019] The second aspect of this invention provides a spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization system, comprising: a transmitting device for the deep-water heterogeneous network, including: a shore-based intelligent clock center; the sparse coding module of the shore-based intelligent clock center is used to perform sparse decomposition on the time-series clock deviation data of each underwater node based on a trained overcomplete dictionary to obtain an original sparse vector, and to compress and sample the original sparse vector using a dynamic chaotic observation matrix generated from the time-series clock deviation data to obtain dynamically compressed observation data; the optical transmitting module of the shore-based intelligent clock center is used to convert the dynamically compressed observation data into an optical signal; an umbilical cable, connected to the optical transmitting module, is used to transmit the optical signal of the transmitting device to the receiving device of the deep-water heterogeneous network; the deep-water heterogeneous network... The receiving device includes: an underwater intelligent routing module, wherein the optical receiving module of the underwater intelligent routing module is connected to the umbilical cable for receiving the optical signal and demodulating the optical signal to obtain an observation vector; the edge computing unit of the underwater intelligent routing module is used to construct a neural network architecture based on a fixed number of layers, using a training method consisting of simulated sparse vectors and noisy chaotic observation vectors to train each layer of the network to generate a deep unfolded network model, and the deep unfolded network model calculates and reconstructs the sparse vector based on the observation vector, and recovers the clock deviation signal from the reconstructed sparse vector through the overcomplete dictionary; the synchronization control module of the underwater intelligent routing module is used to compare the clock deviation signal with a deviation threshold and generate a clock synchronization calibration command to achieve data synchronization of the corresponding underwater node.

[0020] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0021] This invention provides a spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization method and system. It uses a trained overcomplete dictionary to sparsely decompose the timing clock deviation data of each underwater node, and then uses a dynamic chaotic observation matrix generated from the timing clock deviation data to compress and sample the original sparse vector, significantly reducing the amount of data to be transmitted. Furthermore, it transmits optical signals via an umbilical cable, effectively reducing bandwidth usage. The deep unfolded network model is trained end-to-end using a training pair consisting of simulated sparse vectors and noisy chaotic observation vectors. The receiving device inputs the demodulated observation vectors into the trained deep unfolded network model and can directly output reconstructed sparse vectors. This enables accurate signal reconstruction in noisy scenarios, improving signal reconstruction accuracy in low signal-to-noise ratio environments, thereby achieving noise suppression. The deep unfolded network model, combined with an end-to-end training mode and a fixed-layer neural network architecture, solidifies the dynamic iterative process of traditional reconstruction algorithms into deterministic forward propagation steps. This eliminates the need for redundant iterations and real-time parameter adjustments required by traditional reconstruction algorithms. Furthermore, by employing a pre-trained overcomplete dictionary, it can quickly restore reconstructed sparse vectors to clock skew signals, significantly reducing reconstruction time and achieving lower reconstruction latency and faster synchronization response. Finally, by generating clock synchronization calibration commands, it enables efficient data synchronization of corresponding underwater nodes, adapting to the complex transmission environment of heterogeneous deep-water networks and solving the problems of high bandwidth consumption, low signal-to-weight ratio, and poor reconstruction timeliness inherent in traditional network time synchronization technologies. Attached Figure Description

[0022] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other embodiments based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of a spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization method according to an embodiment of the present invention.

[0024] Figure 2 This is a structural block diagram of a spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization system, which is an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of the structure of an electronic device created by the present invention. Detailed Implementation

[0026] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0027] To address the issues of high bandwidth consumption, low signal-to-weight ratio, and poor reconstruction timeliness in network time synchronization technologies, this invention provides a spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization method and system.

[0028] Among them, such as Figure 1 As shown, the spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization method provided by the embodiments of the present invention includes the following steps S1 to S4.

[0029] Step S1: Based on the trained overcomplete dictionary, the timing clock deviation data of each underwater node is sparsely decomposed through the transmitting device of the deep-water heterogeneous network to obtain the original sparse vector.

[0030] Step S2: After converting the dynamic compressed observation data corresponding to the original sparse vector into an optical signal, the transmitting device transmits the optical signal to the receiving device of the deep-water heterogeneous network through an umbilical cable; wherein, the dynamic compressed observation data is obtained by compressing and sampling the original sparse vector through a dynamic chaotic observation matrix generated by time-series clock deviation data.

[0031] Step S3: After the optical signal is demodulated into an observation vector by the receiving device, it is input into the trained deep unfolded network model, and the deep unfolded network model outputs a reconstructed sparse vector. The deep unfolded network model is based on a neural network architecture with a fixed number of layers, and is obtained by end-to-end training of each layer of the network using training pairs composed of simulated sparse vectors and noisy chaotic observation vectors.

[0032] Step S4: By using an overcomplete dictionary, the reconstructed sparse vector is restored to a clock deviation signal, and a clock synchronization calibration command is generated to achieve data synchronization of the corresponding underwater nodes.

[0033] An overcomplete dictionary, in the field of sparse representation, is a redundant set composed of several atoms with signal representation capabilities. Its number of atoms is greater than the dimension of the signal to be represented. It can accurately approximate the target signal through a linear combination of a small number of atoms and is the core foundation for realizing sparse decomposition of signals.

[0034] The atomic features of a general dictionary (generated based on a preset general basis, such as a discrete cosine transform basis and a wavelet basis) are fixed and cannot adapt to the specific characteristics of underwater time-series clock deviation data (such as dynamic fluctuations, environmental interference coupling, and individual node differences). Direct use of such a dictionary leads to low sparse decomposition accuracy and incomplete redundancy removal. Through training, the atomic features can be made to highly match the statistical regularities and dynamic characteristics of the target clock deviation data, thereby improving the efficiency and accuracy of sparse decomposition. This provides a reliable prerequisite for subsequent compressed sampling and signal reconstruction, ensuring the accuracy of clock synchronization.

[0035] Overcomplete dictionaries can be trained using an iterative optimization method based on sample adaptation. The core of this method is to iteratively optimize sparse coefficients and update atoms, gradually adapting the dictionary to the features of the training samples to ultimately meet the accuracy requirements of sparse representation. For example, using a training sample set composed of time-series clock skew data from multiple time periods and underwater nodes in a deep-water heterogeneous network as a foundation, an initial dictionary and sparsity constraints are first set. An optimization algorithm is then used to solve for the sparse coefficients that fit the initial dictionary, ensuring that the combination of these coefficients and the dictionary approximates the training samples while satisfying the sparsity requirements. Then, based on the sample feature information reflected by the solved sparse coefficients, the atoms in the dictionary are updated in a targeted manner to improve the dictionary's ability to represent samples. This iterative process of sparse coefficient optimization and atomic updates is repeated until the sample reconstruction accuracy reaches a preset standard or the iteration reaches a preset upper limit, ultimately resulting in a well-trained overcomplete dictionary that is highly adapted to the features of the time-series clock skew data.

[0036] The transmitting equipment of a deep-sea heterogeneous network is a core control node of the network with core functions such as providing a reference clock reference, sparse data processing, signal conversion and transmission control. Essentially, it is a key device that provides a clock synchronization reference for the entire underwater network and performs data preprocessing and signal transmission, such as a shore-based smart clock center.

[0037] Underwater nodes are various functional terminals deployed in the underwater environment in deep-water heterogeneous networks that need to be synchronized with a reference clock. These include, but are not limited to, underwater sensors (such as temperature, salinity, and depth sensors, and acoustic sensors), underwater communication relay nodes, and underwater execution devices (such as underwater robots and smart moorings).

[0038] Timing clock bias data refers to the time series data formed by continuously collecting and sorting the deviation values ​​between the local clocks of multiple underwater nodes and the unified reference clock of the network (such as the reference clock of the transmitting device) in a deep-water heterogeneous network.

[0039] In step S1, the transmitting device has pre-stored a trained overcomplete dictionary adapted to the characteristics of underwater timing clock deviation data. At the same time, it acquires the continuous high-dimensional timing clock deviation data of each underwater node at the current moment. Then, using the overcomplete dictionary as the basis vector set, it solves the timing clock deviation data of each node using a sparse optimization algorithm (such as L1 regularization-based minimization, matching pursuit algorithms, etc.) to find a coefficient vector. This coefficient vector, when combined with a linear combination of some atoms in the overcomplete dictionary, can accurately approximate the timing clock deviation data to be decomposed within a preset error range. Furthermore, this coefficient vector satisfies the sparsity requirement that the vast majority of its elements are 0 or close to 0. Finally, a qualified original sparse vector is generated for the timing clock deviation data of each underwater node, which is used in the subsequent dynamic compression sampling stage.

[0040] Step S1 of this invention achieves accurate sparse representation of timing clock deviation data through an adaptive overcomplete dictionary. This transforms the high-dimensional redundant original timing data into a low-dimensional sparse vector form, significantly reducing the bandwidth consumption and node energy consumption of underwater networks, thus adapting to underwater resource-constrained scenarios. Furthermore, while removing redundant information, it fully preserves the core characteristics of clock deviation, such as timing change trends, abnormal fluctuations, and individual node differences. This provides reliable data support for the receiving device to subsequently reconstruct the clock deviation signal with high precision and achieve efficient dynamic compression sampling, indirectly ensuring the accuracy and efficiency of clock synchronization in the entire deep-sea heterogeneous network.

[0041] The dynamic chaotic observation matrix is ​​a low-dimensional matrix constructed based on the characteristics of chaotic systems, with a dimension much lower than that of the original sparse vector. The element values ​​of the dynamic chaotic observation matrix change dynamically with the input data, possessing the pseudo-randomness, sensitivity, and low correlation of chaotic signals, making it a core carrier for achieving compressed sampling and data encryption.

[0042] The dynamic chaotic observation matrix of this invention uses the time-series clock deviation data of the current underwater node as the core dynamic seed, extracting specific feature parameters of the data (such as instantaneous deviation peak value, time-series fluctuation variance, and deviation difference between adjacent time points). These parameters are used as initial conditions or control parameters for chaotic mapping (such as Logistic mapping, Lorenz mapping, Chen mapping, etc.). A random number sequence is generated iteratively through chaotic mapping. After normalization and quantization of the sequence, a low-dimensional matrix that meets the requirements of compressed sampling is constructed (the number of rows is the preset sampling dimension, and the number of columns is consistent with the original sparse vector dimension). Because the time-series clock deviation data has real-time dynamic characteristics, the seed parameters input each time are unique, and the generated chaotic observation matrix is ​​also completely different, thus achieving the physical layer dynamic encryption characteristic of "one-time encryption" and preventing data from being maliciously intercepted.

[0043] Dynamically compressed observation data is achieved by linearly projecting the original sparse vectors using a dynamic chaotic observation matrix, significantly reducing the dimensionality of the originally high-dimensional sparse data. During the mapping process, the low correlation of the matrix ensures that core information related to clock skew is not lost after sampling, while its pseudo-randomness enhances the confidentiality of data transmission. Ultimately, this achieves efficient data compression, reduces transmission resource consumption, and guarantees the integrity and security of the sampled data, making it suitable for scenarios with limited resources and high-reliability transmission requirements in deep-sea heterogeneous networks.

[0044] The dynamically compressed observation data is first converted into a standardized binary digital signal through analog-to-digital conversion or quantization encoding. Then, based on the anti-interference requirements, bandwidth limitations, and link characteristics of the deep-water transmission scenario, an appropriate optical modulation technology is selected. The data is loaded by mapping the logic state of the digital signal to the amplitude, frequency, or phase changes of the optical signal. Finally, the optical transmitting device converts the modulated electrical signal into an optical signal adapted to the transmission link, completing a universal and highly reliable conversion from data to optical domain signal.

[0045] Optical modulation techniques include: ASK (Amplitude Shift Keying), FSK (Frequency Shift Keying), QPSK (Quadrature Phase Shift Keying), and DPSK (Differential Phase Shift Keying) modulation techniques.

[0046] Umbilical cables are composite transmission cables designed specifically for deep-water / underwater environments. Their core structure includes optical fibers, power lines, signal lines, and a high-strength protective layer, combining signal transmission, power supply, and physical protection functions.

[0047] The receiving equipment of the deep-water heterogeneous network is the core processing node responsible for optical signal reception, signal recovery and observation data preprocessing. Specifically, it may include underwater aggregation nodes, shore-based data processing centers, underwater communication gateways, underwater environmental monitoring nodes, etc.

[0048] The transmitting device transmits optical signals via an umbilical cable, securely and efficiently transferring dynamically compressed observation information carrying the core characteristics of clock deviation to the receiving device. This provides reliable input for subsequent sparse vector reconstruction, clock deviation signal recovery, and calibration command generation. Leveraging the low loss and high bandwidth of optical signals, as well as the resistance to interference and corrosion in deep-water environments, the device is adaptable to resource-constrained deep-water scenarios while ensuring data transmission integrity and low error rate. This lays the transmission foundation for high-precision clock synchronization in the entire deep-water heterogeneous network.

[0049] In step S2, the transmitting device extracts specific parameters from the timing clock offset data of each underwater node, inputs a chaotic mapping as initial conditions, iteratively generates a pseudo-random number sequence, and performs normalization processing to construct a dynamic chaotic observation matrix that matches the dimension of the original sparse vector. Next, a low-dimensional projection is performed on the high-dimensional original sparse vector to obtain dynamically compressed observation data with a dimension of 1 / 10 to 1 / 5 of the original vector. Then, the dynamically compressed observation data is quantized into binary digital signals and mapped to the phase difference information of adjacent optical pulses using optical modulation technology to obtain an optical signal suitable for transmission. Finally, it is stably transmitted to the receiving device via an umbilical cable.

[0050] In step S2 of this invention embodiment, the original sparse vector is compressed and sampled by a dynamic chaotic observation matrix generated by time-series clock deviation data. The resulting dynamically compressed observation data is converted into an optical signal and transmitted via an umbilical cable, achieving the dual effects of data dimensionality reduction and compression and one-time encryption. This ensures the efficiency, security, and integrity of data transmission and provides reliable support for subsequent sparse vector reconstruction and clock synchronization in deep-water heterogeneous networks for the receiving device.

[0051] The observation vector is a low-dimensional data carrier obtained by the receiving device after demodulating the optical signal. It is the restored form of the dynamically compressed observation data at the receiving device, and its dimension is consistent with that of the dynamically compressed observation data.

[0052] The receiving equipment receives the optical signal transmitted through the umbilical cable via an optical receiving front-end (such as a photodetector). It first converts the optical signal into a corresponding electrical signal, and then, based on the modulation method of the transmitting equipment (such as phase, amplitude, or frequency modulation), uses a demodulation algorithm to restore the digital information in the electrical signal, thus canceling out noise and distortion during transmission. Finally, the demodulated digital signal undergoes standardization processing (such as denoising and quantization calibration) to obtain a low-dimensional vector, i.e., the observation vector, that corresponds one-to-one with the dynamically compressed observation data from the transmitting equipment.

[0053] A simulated sparse vector is a randomly generated vector that satisfies the sparsity requirement.

[0054] The noisy chaotic observation vector is obtained by performing a compression sampling operation on a simulated sparse vector using a chaotic observation matrix, and finally adding Gaussian white noise n to the result.

[0055] The Deep Unfolded Network model is a neural network designed with a fixed number of layers. It completes end-to-end training by simulating training pairs consisting of sparse vectors and noisy chaotic observation vectors. It maps the iterative logic of traditional sparse reconstruction to the optimization of parameters of each layer of the network. After training, it has strong noise resistance, high accuracy and adaptability to chaotic scenarios. It can accurately learn the inverse mapping relationship between observation vectors and sparse vectors. After inputting the observation vector, it can directly restore the core features through the optimized network parameters and output a reconstructed sparse vector that is highly consistent with the original sparse vector.

[0056] In step S3 of this invention embodiment, the receiving device demodulates the optical signal into an observation vector and inputs it into a trained deep unfolding network model. Through the end-to-end training of the model, the model learns the mapping relationship between the observation vector and the sparse vector and its anti-noise capability, thereby achieving efficient and accurate output of the reconstructed sparse vector. This not only cancels noise interference during transmission and sampling but also adapts to the specificity of chaotic compressed sampling, providing highly reliable data support for subsequent clock deviation signal recovery.

[0057] The clock skew signal is a clock skew signal that is consistent with the timing characteristics and skew patterns of the original data, obtained by performing a linear combination operation on the reconstructed sparse vector and the overcomplete dictionary (i.e., using the non-zero coefficients in the sparse vector to call the corresponding atoms in the dictionary and weighted superimpose them).

[0058] Based on the reconstructed clock deviation signal, the clock deviation value (including deviation magnitude and lead / lag direction) of each underwater node is extracted. The corresponding clock adjustment amount is calculated based on the deviation value, and a node identifier and execution timestamp are appended to form a clock synchronization calibration command. This command is then transmitted to the corresponding underwater node via the communication link of the deep-sea heterogeneous network. Upon receiving the command, the node adjusts its local clock according to the instruction to align it with the reference clock, ultimately achieving consistency in the data timestamps of all nodes and completing data synchronization.

[0059] In step S4 of this invention embodiment, the reconstructed sparse vector is restored to a precise clock deviation signal using a trained overcomplete dictionary. Based on this, a targeted clock synchronization calibration command is generated and sent to the corresponding underwater node, thereby aligning the node's local clock with the reference clock. This ensures the consistency of timestamps for each node's data and completes the data synchronization of the deep-sea heterogeneous network.

[0060] The spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization method provided by the embodiments of the present invention performs sparse decomposition of the timing clock deviation data of each underwater node through a trained overcomplete dictionary, and then uses the dynamic chaotic observation matrix generated by the timing clock deviation data to compress and sample the original sparse vector, which greatly reduces the amount of data to be transmitted, and effectively reduces bandwidth occupation by transmitting optical signals through umbilical cables.

[0061] The deep unfolded network model is trained end-to-end using training pairs consisting of simulated sparse vectors and noisy chaotic observation vectors. After the receiving device inputs the demodulated observation vectors into the trained deep unfolded network model, it can directly output the reconstructed sparse vectors. This enables accurate signal reconstruction in noisy scenarios, improves signal reconstruction accuracy in low signal-to-noise ratio environments, and thus achieves noise suppression.

[0062] The deep unfolded network model, combined with an end-to-end training mode and a fixed-layer neural network architecture, solidifies the dynamic iterative process of traditional reconstruction algorithms into deterministic forward propagation steps. This eliminates the need for redundant iterations and real-time parameter adjustments required by traditional reconstruction algorithms. Furthermore, by employing a pre-trained overcomplete dictionary, it can quickly recover clock skew signals from reconstructed sparse vectors, significantly reducing reconstruction time and achieving lower reconstruction latency and faster synchronization response. Finally, by generating clock synchronization calibration commands, it enables efficient data synchronization of corresponding underwater nodes, adapting to the complex transmission environment of heterogeneous deep-sea networks.

[0063] Furthermore, the method provided in the embodiments of the present invention preferably includes the following steps before step S1:

[0064] Step S01: Based on the randomly generated initial dictionary, calculate the sparse coefficient matrix that minimizes the reconstruction error between the reconstructed sample matrix and the training sample set and satisfies the sparsity constraint; wherein, the training sample set includes the time-series clock deviation data of multiple underwater nodes in multiple time periods of the deep-water heterogeneous network; the reconstructed sample matrix is ​​obtained by multiplying the initial dictionary with the sparse coefficient matrix.

[0065] Step S02: Based on the non-zero coefficients of the sparse coefficient matrix, determine the sample index corresponding to the initial dictionary and construct the restricted residual matrix corresponding to the sample index.

[0066] Step S03: Based on the new atoms and coefficients obtained by performing singular value decomposition on the restricted residual matrix, iteratively update the initial dictionary and sparse coefficient matrix until the change in reconstruction error is less than a preset threshold or the maximum number of iterations is reached, thus obtaining the trained overcomplete dictionary.

[0067] Specifically, during the initialization phase or offline state of the deep-water heterogeneous network data synchronization system, the collected data... Within a time period Clock offset data for each underwater node (e.g., one sampling point every 5 minutes), the clock offset can be obtained by comparing with a high-precision reference clock. The clock offset at each time point... The data of each node is arranged into a column vector. ,in, Indicates the first A column vector of clock skew data over a time period It is an index of M time periods. , It is the space of real numbers. express Dimensional specifications: The corresponding number of underwater nodes, 1 represents It is a single column vector. The training sample set is... .

[0068] Furthermore, regarding Each row (i.e., the time series data of each node) undergoes mean removal and normalization to eliminate the influence of DC components and dimensions, making the data more suitable for training.

[0069] The deep-sea heterogeneous network data synchronization system is an integrated technical system used to perform the entire process of steps S1 to S4 mentioned above. Its core objective is to achieve nanosecond-level high-precision time synchronization of multiple underwater nodes. It integrates offline feature learning, online data processing, secure compressed transmission, rapid intelligent reconstruction, and synchronization calibration.

[0070] Initialize dictionary The Orthogonal Matching Pursuit (OMP) algorithm is used to solve for the sample set under the current dictionary. sparse coefficient matrix Specifically, the preferred formula for solving the sparse coefficient matrix is:

[0071] ;in It is a complete dictionary to be learned ( ), each of its columns It is called an atom. It is a sparse coefficient matrix. It is the Frobenius norm. Let L0 norm be denoted as and let vector be denoted as . The number of non-zero elements in the neutron. It is a preset sparsity constraint, that is, each signal can be used at most... Represented by one atom.

[0072] In overcomplete dictionary training, for a specific column of an atom in the initial dictionary, the row corresponding to that atom in the sparse coefficient matrix is ​​first found. The column indices of the non-zero coefficients in that row are then selected, yielding the sample indices that use that atom. These are ultimately aggregated into a sample set, providing a precise sample range for subsequent atom updates and residual calculations. All sparse coefficient vectors... The Middle A set of indices of non-zero elements It can be represented as: .

[0073] After determining the sample index, the error information generated by the sample corresponding to the index is screened from the full error data source (such as static residual matrix, dynamic error sequence, joint error field, etc.), the interference error of irrelevant sample is removed, and finally the screened error subset is normalized and integrated according to preset dimensions (such as sample, feature, time dimension, etc.) to form a restricted residual matrix that focuses only on the effective error related to the target optimization object.

[0074] Singular value decomposition is used to mathematically optimize the constrained residual matrix, extracting new atoms and corresponding new sparse coefficients that minimize the residual, thus achieving synergistic optimization of atoms and coefficients. Subsequently, K-SVD decomposition is employed to update each column of the initial dictionary with the new atoms, and the corresponding old values ​​in the sparse coefficient matrix are replaced with the new coefficients, completing a single iteration update.

[0075] Finally, the process of iteratively repeating sample indexing, restricting residual matrix construction, and singular value decomposition optimization updates continues until the change in reconstruction error meets the preset convergence condition (less than a threshold or reaching the maximum number of iterations). This ultimately yields a well-trained, comprehensive dictionary with strong adaptability that efficiently represents underwater clock deviation signals, which can be represented as: ( ).

[0076] The trained, complete dictionary is pre-installed on the transmitting and receiving devices of the deep-water heterogeneous network.

[0077] The method provided in the embodiments of this invention, based on a randomly generated initial dictionary, solves for the sparse coefficient matrix that minimizes the reconstruction error between the reconstructed sample matrix obtained by multiplying the initial dictionary and the sparse coefficient matrix and the training sample set composed of time-series clock deviation data from multiple underwater nodes over multiple time periods, under sparsity constraints. Then, the sample indices corresponding to each atom in the initial dictionary are locked by the non-zero coefficients of the sparse coefficient matrix, and a constraint residual matrix focusing on the relevant sample errors is constructed. After obtaining new atoms and new coefficients through singular value decomposition, the initial dictionary and the sparse coefficient matrix are iteratively updated until the reconstruction error meets the convergence condition. The resulting overcomplete dictionary can accurately fit the nonlinear characteristics of underwater clock drift and the patterns of multi-node time-series data, ensuring the effectiveness of the sparse representation of the signal and reducing the reconstruction error through precise error focusing and iterative optimization, laying the foundation for subsequent efficient sparse decomposition.

[0078] Furthermore, in step S02, the restricted residual matrix corresponding to the sample index is constructed, including the following steps: based on the initial dictionary, the residual matrix corresponding to each column atom is constructed; wherein, the residual matrix is ​​constructed based on the difference between the training sample set and the corresponding reconstruction matrix; the reconstruction matrix is ​​the sum of multiplying the atoms in other columns except the current column atom with the corresponding coefficient row vector in the sparse coefficient matrix; based on the sample index, the corresponding restricted residual matrix is ​​selected from each residual matrix.

[0079] Specifically, when constructing the residual matrix corresponding to each column of atoms based on the initial dictionary, each column of atoms in the initial dictionary needs to be processed one by one: For the current column of atoms to be processed, first select all other columns of atoms in the initial dictionary except for this column, then find the coefficient row vectors in the sparse coefficient matrix that match these other columns of atoms one by one, multiply each other column of atoms by its corresponding coefficient row vector, and then sum all the product results to obtain the reconstruction matrix corresponding to the current atom. Then, subtract this reconstruction matrix from the training sample set, and the result is the residual matrix corresponding to the current column of atoms.

[0080] Furthermore, the preferred formula for calculating the residual matrix is: ,in Represents the sparse coefficient matrix The OK, It is to remove atoms After the contribution, the approximation error of all samples. Indicates the removal of the first The reconstruction matrix corresponding to the column atoms.

[0081] When filtering the restricted residual matrix based on sample indices, the sample indices are the sample identifiers previously associated with the atoms in the current column. By simply retaining only the portions corresponding to these sample indices from the residual matrix corresponding to the current atom and removing the contents corresponding to samples unrelated to that atom, the restricted residual matrix corresponding to the current column atom can be obtained. For example, the... The restricted residual matrix corresponding to the sample indices of non-zero elements is: .

[0082] The method described above, provided by the embodiments of the present invention, constructs the corresponding residual matrix for each atom in the initial dictionary. This operation can accurately separate the independent contribution error of each atom to the reconstruction of the training samples, avoid the confusion of errors of different atoms, and clearly quantify the actual influence of each atom.

[0083] The reconstruction matrix retains only the parts other than the current atom, further ensuring that the residual matrix is ​​the approximation error specific to the current atom, providing a precise target basis for the optimization of subsequent atoms.

[0084] By combining the sample index with the restricted residual matrix, we can eliminate the interference of sample error that is irrelevant to the current atom and retain only the effective error information that is truly related to the atom. This makes the subsequent dictionary atom updates based on these matrices more targeted and effectively improves the accuracy and efficiency of overcomplete dictionary training.

[0085] Furthermore, in step S03 of the present invention, the initial dictionary and sparse coefficient matrix are iteratively updated based on the new atoms and new coefficients obtained by performing singular value decomposition on the restricted residual matrix. This includes the following steps: performing singular value decomposition on the restricted residual matrix to obtain the corresponding left singular vector matrix, diagonal matrix, and right singular matrix; generating corresponding new atoms and new coefficients based on the left singular vector matrix, diagonal matrix, and right singular matrix; wherein, the new atom is the first column of the left singular vector matrix; and the new coefficient is the product of the first element of the diagonal matrix and the first column of the right singular matrix.

[0086] Specifically, when performing singular value decomposition on the restricted residual matrix, we obtain a left singular vector matrix containing the core feature directions of the residual data, a diagonal matrix of singular values ​​arranged in descending order of contribution, and a right singular matrix corresponding to the coefficient transformation directions. The first column of the left singular vector matrix is ​​selected as the new atom (because it corresponds to the feature direction with the largest contribution in the residual), and the first largest singular value in the diagonal matrix is ​​multiplied by the first column of the right singular matrix to obtain new coefficients that fit the new atom.

[0087] Furthermore, regarding the restricted residual matrix The expression for performing Singular Value Decomposition (SVD) is: ,in, Describes a left singular vector matrix. Represents a diagonal matrix. This represents the transpose of a right singular matrix.

[0088] Will The first column will be used as the new atom. The first element in the multiplier The first column as for The new coefficients are updated column by column using the above method to update the atoms in the initial dictionary and simultaneously update the sparse coefficient matrix, thereby achieving iterative updates of the initial dictionary and the sparse coefficient matrix.

[0089] The method provided by the embodiments of the present invention focuses on the most critical feature information in the residual matrix, enabling new atoms to accurately match the error patterns of the current associated samples and minimizing the reconstruction error corresponding to the atom. At the same time, the matching degree between the new coefficients and the new atoms is optimized synchronously, making the atom update more targeted. This effectively improves the accuracy and convergence efficiency of the overcomplete dictionary iteration, helping the dictionary to adapt to the sparse representation requirements of underwater clock deviation signals more quickly.

[0090] Further, in step S1, the current data to be synchronized is obtained. Time deviation data from several underwater nodes are used to construct the original signal vector. The vector index corresponds to the node's physical ID, and the numerical value corresponds to the time deviation quantization value.

[0091] Using a trained overcomplete dictionary ,right Perform sparse decomposition to obtain highly sparse original sparse vectors. (Right now ),at this time Its sparsity is far superior to that of the original signal.

[0092] Furthermore, in step S2 of this embodiment, converting the dynamically compressed observation data corresponding to the original sparse vector into an optical signal preferably includes the following steps: generating an initial chaotic value based on the current timestamp of the time-series clock deviation data through a unidirectional irreversible numerical mapping operation; wherein the initial value is in the range of 0 to 1; inputting the initial value into the chaotic mapping equation for iterative calculation to generate a dynamic chaotic observation matrix that satisfies a preset length; compressing and sampling the original sparse vector based on the dynamic chaotic observation matrix to obtain the dynamically compressed observation data; and converting the dynamically compressed observation data into an optical signal based on the mapping relationship between the electrical signal and the phase difference of adjacent optical pulses.

[0093] Specifically, the one-way irreversible numerical mapping operation includes cryptographic hash functions and FNV (Fowler-Noll-Vo) hash functions, preferably cryptographic hash functions (such as SHA-56). This operation possesses one-way irreversibility, collision resistance, and input sensitivity, ensuring that different timestamps correspond to unique mapping results and preventing the original timestamp from being deduced from the mapping result. This guarantees the dynamic randomness of the chaotic initial value and provides underlying security support for subsequent data encryption. The timestamp of the current frame is calculated using the SHA-256 hash function to obtain a fixed-length (256-bit) hash value. A specified bit segment of this hash value (such as the high 32 bits or the low 64 bits) is extracted and normalized, mapping it to the interval (0, 1) to obtain the chaotic initial value, ensuring the uniform distribution and unpredictability of the initial value.

[0094] Chaotic mapping equations include: one-dimensional or high-dimensional chaotic mapping models, preferably Logistic chaotic mapping equations, whose expressions are as follows: ,in, This represents the output value of the (n+1)th iteration. This represents the output value of the nth iteration. Indicates system parameters. Sets system parameters. It is in a state of complete chaos.

[0095] Input the initial value into the above Logistic chaotic mapping equation for iterative operation. The number of iterations is the same as the number of columns of the preset dynamic chaotic observation matrix. The output value of each iteration is used as a row element of the observation matrix. Repeat the above process until a dynamic chaotic observation matrix with M rows and K columns is generated (where M << K, M is the dimension of the observation vector, and K is the dimension of the original sparse vector. This chaotic observation matrix is dynamically updated with the time stamp, which not only meets the sparse signal sampling requirements of compressive sensing but also encrypts the observation process through the chaotic characteristics, making it impossible for unauthorized parties to reverse the original signal from the intercepted observation data.

[0096] Based on the dynamic chaotic observation matrix, perform a linear projection calculation on the original sparse vector, and the -dimensional sparse coefficient vector can be compressed into -dimensional observation vector, obtaining the dynamically compressed observation data, while achieving data encryption and dimensionality reduction.

[0097] The expression of the linear projection operation is preferably: , where represents the dynamically compressed observation data, represents the dynamic chaotic observation matrix, represents the original sparse vector. This linear projection process destroys the correlation of the original signal through the randomness of the chaotic observation matrix, making the compressed observation data have both dimensionality reduction efficiency and encryption security, and solving the problem that data is easily cracked in the traditional compressive sampling process.

[0098] Optical modulation technologies based on the mapping relationship between electrical signals and the phase difference of adjacent optical pulses include: differential phase shift keying (DPSK), differential quadrature phase shift keying, and coherent differential phase shift keying.

[0099] Convert the numerical characteristics (such as amplitude size, sign change, quantization level, etc.) of the electrical signal (analog electrical signal or digital electrical signal) corresponding to the dynamically compressed observation data into the relative phase difference between continuous optical pulses through a preset mapping rule, and use the phase difference as the information carrier to achieve optical signal modulation, rather than relying on the absolute phase or intensity of the optical pulse to carry information. The advantage of this mapping relationship is that the phase difference is less affected by fiber dispersion and intensity noise during long-distance transmission, which can significantly improve the reliability of underwater fiber optic transmission and adapt to the long-distance secure transmission requirements of dynamically compressed observation data (including encrypted information).

[0100] This embodiment of the invention preferably adopts differential phase shift keying modulation technology. First, each vector element of the dynamically compressed observation data ( = 1, 2,..., M) is converted into the corresponding electrical signal symbol. For the mth electrical signal symbol, its corresponding target phase difference is In actual transmission, the absolute phase of the symbol is not directly transmitted. ( Instead, it transmits the phase difference between itself and the previous light pulse. (in, (This refers to the phase of the (m-1)th optical pulse).

[0101] By adjusting the phase of the input optical carrier using a Mach-Zehnder modulator (MZM), the adjacent phase differences of the output optical pulses correspond one-to-one with the electrical signal symbols, thus embedding the numerical information of the dynamically compressed observation data into the phase differences of the optical signal. For example, when the dynamically compressed observation data is a binary quantization result, the phase difference corresponding to electrical signal "0" and the phase difference corresponding to electrical signal "1" can be set to achieve DPSK modulation of binary information; if it is a multi-level quantization result, the phase difference can be set to 0, 2 / ... , Multiple levels are available to accommodate higher capacity data transmission.

[0102] The method provided by the embodiments of the present invention generates a chaotic initial value in the 0-1 interval by mapping the current timestamp associated with the timing clock deviation to a one-way irreversible mapping. A dynamic observation matrix is ​​constructed through chaotic mapping iteration. While compressing and sampling the original sparse vector to reduce its dimensionality, the chaotic characteristics are used to give the data encryption security. Finally, the optical signal conversion is completed by mapping the phase difference between the electrical signal and the adjacent optical pulse. This not only ensures the resistance to intensity noise during data transmission, but also achieves the synergistic effect of efficient compression, secure encryption and reliable transmission of sparse data.

[0103] Furthermore, in the embodiments of the present invention, the initial value is input into the chaotic mapping equation for iterative calculation to generate a dynamic chaotic observation matrix that satisfies a preset length. Preferably, this includes the following steps: inputting the initial value into the chaotic mapping equation for iterative calculation, discarding the transient values ​​generated by the previous set number of iterations, and generating a pseudo-random sequence of a preset length; rearranging the purified pseudo-random sequence into an initial matrix of a preset dimension according to row priority or column priority rules; normalizing the column vectors of the initial matrix, and then optimizing the orthogonality of the column vectors through vector projection decomposition to obtain the final dynamic chaotic observation matrix.

[0104] Specifically, the initial value is input into the chaotic mapping equation for iterative calculation, and the length of the generated pseudo-random sequence is: ,in, Indicates the length of the instantaneous value. This represents the total length of the pseudo-random sequence.

[0105] Discard the first few lines of the pseudo-random sequence The instantaneous values, such as those from the first 100 iterations, can eliminate transient effects and ensure that the chaotic properties of the sequence are specific.

[0106] The following The values ​​are rearranged in row-major or column-major order. initial matrix ( Next, regarding Perform column normalization and vector projection decomposition to orthogonalize the column vectors, resulting in the final observation matrix. .

[0107] The vector projection decomposition for achieving column vector orthogonality optimization includes Gram-Schmidt orthogonalization, Householder Transformation orthogonalization, and Givens Rotation orthogonalization. The preferred embodiment of this invention utilizes Gram-Schmidt orthogonalization, which simultaneously achieves four objectives—optimized observation matrix sampling performance, enhanced encryption security, dynamic scene adaptation, and reliable transmission support—under constraints of low complexity and low energy consumption, thus meeting the core requirements of efficient compression, secure encryption, and reliable transmission of sparse data.

[0108] Furthermore, the method of this invention, prior to step S3, preferably includes the following steps: constructing a deep unfolded network model based on a neural network architecture with a fixed number of layers; wherein the operation rules of the deep unfolded network model include: fusing the received observation information with the reconstruction results of the previous layer based on a learnable weight matrix, and then obtaining the reconstruction results of each layer of the network through soft thresholding; and performing end-to-end training on each weight matrix and the threshold parameters corresponding to the soft thresholding calculation based on training pairs composed of randomly generated simulated sparse vectors that satisfy sparsity constraints and noisy chaotic observation vectors, combined with a loss function, to obtain a trained deep unfolded network model.

[0109] Specifically, the deep unfolded network model is a neural network based on the LISTA (Learned Iterative Shrinkage-Thresholding Algorithm) concept. It has a fixed number of layers, L, which can be set according to the compressed sensing iterative reconstruction logic to adapt to the dimensionality ratio of the original sparse vector and the observation vector, avoiding computational redundancy caused by excessively deep networks and adapting to resource-constrained scenarios. This network unfolds the single iteration of the traditional ISTA algorithm into a single layer of the neural network.

[0110] The operational rules of the deep unfolded network model are as follows: two types of learnable weight matrices are used to perform dimension matching linear transformation on the received noisy chaotic observation information and feature enhancement on the previous layer reconstruction results. The two are then fused and input into the soft threshold calculation module to remove small-amplitude redundant components in the fused features and enhance sparsity, thereby obtaining the reconstruction results of each layer of the network (when there are no previous reconstruction results in the first layer, only initial processing is performed based on the observation information).

[0111] During the training phase of the deep unfolded network model, simulated sparse vectors that satisfy preset sparsity constraints (consistent with the sparsity characteristics of real-world scene data) are randomly generated. These vectors are then linearly projected onto a dynamic chaotic observation matrix and superimposed with noise from underwater optical transmission scene matching (such as Gaussian white noise or fiber intensity noise) to generate noisy chaotic observation vectors, forming a massive number of training pairs. A loss function (such as mean squared error, or a combination of mean squared indifference and sparsity regularization) is used to perform end-to-end joint training on all learnable weight matrices and independently configured soft threshold parameters for each layer. The parameters are continuously optimized to minimize reconstruction error and sparsity bias, ultimately resulting in a well-trained deep unfolded network model.

[0112] Specifically, the training pair can be represented as ,in, It is a randomly generated simulated sparse vector that satisfies sparsity constraints. It is a noisy, chaotic observation vector with added Gaussian white noise. , where n is the added Gaussian white noise to simulate channel error.

[0113] The preferred expression for the loss function is: ,in, Represents the loss function. Indicates batch size, This represents the simulated sparse vector corresponding to the p-th sample. This represents the reconstructed sparse vector output by the deep unfolded network model corresponding to the p-th sample. This represents the square of the L2 norm.

[0114] The method provided by the embodiments of this invention adapts to resource-constrained scenarios through a fixed-layer architecture, fuses observation information with previous-layer reconstruction results using learnable weights, enhances sparsity by combining soft threshold calculation, and then uses simulated vectors that satisfy sparsity constraints and noisy chaotic observation vectors to form training pairs. All weights and threshold parameters are optimized end-to-end through a loss function. This not only ensures the accuracy and noise resistance of sparse vector reconstruction, but also achieves synergy between efficient reconstruction and efficient resource utilization by simplifying the architecture and adapting to the characteristics of chaotic observations through targeted training.

[0115] Furthermore, the expression for the operational rules of the k-th layer of the deep unfolded network is:

[0116] .

[0117] in, This represents the output of the (k+1)th layer of the network; This represents the output of the k-th layer network. This represents the observation information received by the depth-unfolded network; This represents the first weight matrix corresponding to the observation information of the k-th layer network; This represents the second weight matrix corresponding to the output of the k-th layer of the network; This represents the soft-threshold activation function value of the k-th layer network. , This represents the input variable for the soft-threshold activation function; Represents a symbolic function; This represents the threshold parameter of the k-th layer network; This represents the maximum value function.

[0118] Subsequently, the parameters of all layers were trained end-to-end using gradient descent. .

[0119] The expression provided in the embodiments of this invention first performs a linear transformation on the input observation information through the observation weight matrix of the k-th layer, and simultaneously enhances the features of the reconstruction result of the previous layer through the output weight matrix of this layer. After linearly fusing these two types of features, they are input into the parameterized soft threshold function of this layer. The direction of effective features is preserved through the sign function, and small-amplitude redundant components are removed through threshold truncation, finally obtaining the reconstruction output of the next layer. Its physical essence is that in each layer iteration, the original observation information and the preliminary reconstruction result of the previous layer are dynamically combined to gradually select effective components that meet the sparsity characteristics, thereby achieving a layer-by-layer approximation of the original sparse vector.

[0120] Furthermore, , This expression represents the initialization rule for the two weight matrices of the k-th layer of a deep unfolded network. "Represents the assignment relationship: the chaotic observation matrix" Transpose divided by the total number of network layers , as the observation weight matrix of the k-th layer The initial value of the identity matrix. minus" transpose and product divided by The result is used as the output weight matrix of the k-th layer. The initial value is determined by combining the prior information of the chaotic observation matrix in compressed sensing to give the weight matrix an initial foundation that fits the sparse vector reconstruction logic, so that the weights are adapted to the fusion requirements of observation information and previous layer reconstruction results from the early stage of training, rather than being randomly initialized.

[0121] The expression provided in the embodiments of this invention, through a learnable design of observation and output dual-weight matrices, allows the observation information and the reconstruction results of the previous layer to be dynamically adapted and fused according to the actual scenario (such as the randomness of chaotic observations and fluctuations in channel noise), avoiding the rigidity of fixed weights. At the same time, the threshold parameters of the soft threshold function are configured independently by layer, improving the flexibility of sparsity control at different iteration stages, which not only strengthens the sparsity characteristics of the reconstruction results but also retains more effective information. This layer-by-layer iterative optimization operation logic allows the reconstruction results to gradually approach the real sparse vector as the number of network layers increases, effectively improving the reconstruction accuracy of sparse vectors. At the same time, it adapts to the fixed-layer architecture of deep unfolded networks, balancing computational efficiency and resource consumption, and is suitable for resource-constrained application scenarios.

[0122] Step S3 of this embodiment of the invention involves receiving the observation vector. Input the trained LISTA network, and then... Layer forward propagation directly outputs the reconstructed sparse coefficient vector. This process involves only matrix multiplication and simple nonlinear activation, and can be completed extremely quickly on FPGAs or dedicated AI chips.

[0123] Furthermore, step S4 of the present invention preferably includes the following steps: reconstructing the sparse vector into a clock bias signal using an overcomplete dictionary; comparing the clock bias signal with a bias threshold, and generating a clock synchronization calibration command when the absolute value of the clock bias signal is greater than the bias threshold; wherein the clock synchronization calibration command includes: the identifier of the target underwater node and the calibration value; and sending the clock synchronization calibration command to the corresponding target underwater node via the underwater local area network through the receiving device to achieve data synchronization of the corresponding underwater node.

[0124] Furthermore, using an overcomplete dictionary containing the characteristic atomic basis of the clock deviation signal, the reconstructed sparse vector is mapped back to the original clock deviation signal through inverse sparse representation. Based on a preset deviation threshold, it is determined whether the absolute value of the signal exceeds the threshold. If it does, a clock synchronization calibration command containing a unique identifier of the target underwater node and the actual calibration value calculated based on the current clock deviation signal is generated. With the help of an underwater local area network adapted to the underwater communication environment, the receiving device sends the command to the corresponding target underwater node, so that the node adjusts its local clock according to the calibration value, and finally realizes the time alignment between the underwater node and the system, and completes data synchronization.

[0125] Specifically, using a dictionary Restore the original time deviation signal: .

[0126] right Perform threshold judgment and set a threshold. traversal Each element ,if Then determine the node There is a time deviation that needs to be calibrated; the deviation value is... Conversely, determine the node. Time is synchronized; no calibration is required.

[0127] The synchronization controller of the underwater intelligent routing module generates a precise synchronization command containing the target node ID and calibration value based on the decision result, and distributes it to the corresponding underwater nodes through an underwater local area network (such as an optical or acoustic communication link).

[0128] The embodiments of this invention also provide a spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization system, which can be applied to the spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization method.

[0129] Among them, such as Figure 2 As shown, the embodiment of the present invention provides a spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization system, comprising: a transmitting end device for the deep-water heterogeneous network, an umbilical cable, and a receiving end device for the deep-water heterogeneous network.

[0130] The transmitting equipment of the deep-water heterogeneous network includes: a shore-based smart clock center, whose sparse coding module is used to perform sparse decomposition on the time-series clock deviation data of each underwater node based on a trained overcomplete dictionary to obtain the original sparse vector, and to compress and sample the original sparse vector using the dynamic chaotic observation matrix generated by the time-series clock deviation data to obtain dynamic compressed observation data; and the optical emission module of the shore-based smart clock center is used to convert the dynamic compressed observation data into optical signals.

[0131] The umbilical cable, connected to the optical transmitter module, is used to transmit optical signals from the transmitting device to the receiving device in the deep-sea heterogeneous network.

[0132] The receiving device of the deep-sea heterogeneous network includes: an underwater intelligent routing module, whose optical receiving module is connected to an umbilical cable to receive optical signals and demodulate them to obtain observation vectors; an edge computing unit of the underwater intelligent routing module to construct a neural network architecture based on a fixed number of layers, using simulated sparse vectors and noisy chaotic observation vectors to train each layer of the network to generate a deep unfolded network model, and using the deep unfolded network model to calculate and reconstruct sparse vectors based on the observation vectors, and then recovering the clock deviation signal from the reconstructed sparse vectors through an overcomplete dictionary; and a synchronization control module of the underwater intelligent routing module to compare the clock deviation signal with a deviation threshold and generate clock synchronization calibration commands to achieve data synchronization of the corresponding underwater nodes.

[0133] The system provided by the embodiments of the present invention uses a sparse coding module of a shore-based smart clock center to perform sparse decomposition on the timing clock deviation data of each underwater node using a trained overcomplete dictionary to obtain the original sparse vector. It then uses a dynamic chaotic observation matrix generated by the timing clock deviation data to perform compressed sampling to obtain dynamic compressed observation data. The optical transmission module then converts the data into an optical signal and transmits it to the receiving device via an umbilical cable.

[0134] The underwater intelligent routing module of the receiving device demodulates the optical signal to obtain the observation vector. The edge computing unit trains the deep unfolded network model based on a fixed-layer neural network architecture, which consists of simulated sparse vectors and noisy chaotic observation vectors. It then calculates and reconstructs the sparse vectors and recovers the clock deviation signal through an overcomplete dictionary.

[0135] The synchronization control module compares the clock deviation signal with the deviation threshold and generates a clock synchronization calibration command. It not only fully incorporates the core technical details of sparse coding, dynamic chaotic compression sampling, and deep unfolding reconstruction in the previous method, but also has the technical effects of efficient data transmission, accurate clock deviation recovery, and reliable underwater node data synchronization, making it suitable for the application scenarios of deep-sea heterogeneous networks.

[0136] Furthermore, the shore-based smart clock center also includes an atomic clock source to provide a highly stable 10MHz frequency signal and a 1PPS (pulses per second) signal as an absolute time reference.

[0137] The shore-based smart clock center also includes a data acquisition module, which is connected to the atomic clock source and each underwater node. It is used to compare the local clock of each underwater node with the reference clock of the atomic clock source and collect the timing clock deviation data of each underwater node.

[0138] The sparse coding module is preferably a sparse coding engine (FPGA), which integrates a dictionary storage unit and a chaotic sequence generator internally. The dictionary storage unit stores the pre-trained dictionary in the form of a lookup table. The chaotic sequence generator implements Logistic mapping and hash function to generate observation matrix in real time. .

[0139] Furthermore, the optical emission module preferably uses a narrow linewidth laser (to ensure coherence) to drive a Mach-Zehnder modulator (MZM), and the bias voltage and RF drive signal of the MZM are precisely controlled according to the DPSK coding scheme.

[0140] Umbilical cable, also known as optoelectronic composite umbilical cable, consists of single-mode optical fiber (for data transmission), power lines, and an armored protective layer.

[0141] The optical receiver module, comprising a coherent receiver (local oscillator laser, 90° optical mixer, balanced detector), demodulates the DPSK optical signal into an electrical signal, which is then sampled into a digital signal by a high-speed ADC, i.e., for observation. .

[0142] The edge computing unit, equipped with a high-performance, low-power processor (such as NVIDIA Jetson Orin NX), runs an optimized LISTA neural network model. The model parameters have been quantized to INT8 or FP16 format to improve inference speed.

[0143] The synchronization control module, typically an MCU or FPGA, executes signal decision logic and manages the communication protocol stack of the underwater local area network to ensure reliable distribution of synchronization commands.

[0144] Underwater environment monitoring nodes, or underwater nodes for short, have a built-in local clock, such as a TCXO (Temperature Compensated Crystal Oscillator), which has the function of receiving synchronization commands and calibrating the local clock.

[0145] The system provided in this invention preferably uses a Xilinx Ultrascal processor combined with an FPGA (Field-Programmable Gate Array) as the core processor of the shore-based smart clock center, responsible for dictionary storage, chaotic sequence generation, and compression operations. The optical emission section uses a narrow-linewidth laser with a linewidth of less than 10kHz in conjunction with a 40GHz bandwidth lithium niobate modulator.

[0146] The underwater intelligent routing module integrates NVIDIA Jetson series embedded AI modules and deploys a quantized deep unfolded network model to be responsible for high-speed signal reconstruction.

[0147] Furthermore, in order to verify the effectiveness of step S4, this embodiment details the network training process.

[0148] 1. Dataset Construction: Generate 100,000 sets of simulated underwater clock drift sparse signals as label data. Using a dictionary and randomly generated chaotic matrix Generate corresponding observations , constitute training pairs .

[0149] 2. Network Settings: Configure the number of LISTA network layers. Compared to the traditional ISTA algorithm, which typically requires more than 100 iterations to converge, a 12-layer deep unfolded network can achieve the same accuracy.

[0150] 3. Training strategy: The Adam optimizer is used, and the initial learning rate is set to... The gradient decays every 50 epochs. An incremental training strategy is introduced: first train the first layer, then train the second layer with fixed parameters, and finally perform global fine-tuning to accelerate convergence and avoid gradient vanishing.

[0151] 4. Transfer the trained network parameters (weight matrix) and threshold The quantization is in INT8 format and deployed in an NVIDIA Jetson module underwater. The actual measured inference time for a single inference is less than 150μs.

[0152] In summary, the spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization method and system provided by this invention can achieve the following technical effects:

[0153] (1) By introducing K-SVD dictionary learning, the sparse representation is more in line with the physical characteristics of underwater clock drift. Compared with general bases (such as DCT / wavelet), the reconstruction accuracy is improved by more than 30% under the same compression ratio; or the bandwidth usage is further reduced under the same accuracy.

[0154] (2) The physical layer dynamic encryption of “one-time pad” was realized through the chaotic observation matrix to prevent data from being maliciously intercepted; combined with optical domain DPSK modulation, it effectively resisted channel attenuation and noise in deep water long-distance transmission.

[0155] (3) A deep unfolded network model is adopted, which transforms the complex matrix operations of traditional iterative algorithms into forward inference of a neural network with a fixed number of layers. On edge computing devices, the reconstruction time is reduced from milliseconds to microseconds, perfectly meeting the extremely low latency requirements of deep-sea industrial control.

[0156] An embodiment of the present invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform the method of the embodiment of the present invention.

[0157] refer to Figure 3 The present invention will now describe a structural block diagram of an electronic device that can serve as an embodiment of the present invention, serving as an example of a hardware device applicable to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0158] like Figure 3 As shown, the electronic device includes a computing unit 501, which can perform various appropriate actions and processes based on a computer program stored in ROM (Read-Only Memory) 502 or loaded from storage unit 508 into RAM (Random Access Memory) 503. RAM 503 can also store various programs and data required for the operation of the electronic device. The computing unit 501, ROM 502, and RAM 503 are interconnected via bus 504. An I / O interface (Input / Output Interface) 505 is also connected to bus 504.

[0159] Multiple components in the electronic device are connected to I / O interface 505, including: input unit 506, output unit 507, storage unit 508, and communication unit 509. Input unit 506 can be any type of device capable of inputting information into the electronic device. Input unit 506 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 507 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 508 may include, but is not limited to, disks and optical discs. Communication unit 509 allows the electronic device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, and / or wireless communication transceivers, such as Bluetooth devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0160] The computing unit 501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a CPU (Central Processing Unit), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention can be implemented as a computer program tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via ROM 502 and / or communication unit 509. In some embodiments, the computing unit 501 can be configured to perform the methods described above by any other suitable means (e.g., by means of firmware).

[0161] Computer programs for implementing the methods of embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0162] In the context of embodiments of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, or infrared systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0163] It should be noted that the term "comprising" and its variations used in the embodiments of this invention are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "a plurality" mentioned in the embodiments of this invention are illustrative and not restrictive, and those skilled in the art should understand that unless explicitly indicated otherwise in the context, they should be understood as "one or more".

[0164] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this invention are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0165] The steps described in the method embodiments provided by the present invention can be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of protection of the present invention is not limited in this respect.

[0166] The term "embodiment" in this specification refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily imply the same embodiment, nor does it imply independence or alternativeity from other embodiments. The various embodiments in this specification are described in a related manner, with reference to each other for similar or identical parts. In particular, for apparatus, device, and system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, and relevant details are referred to in the description of the method embodiments.

[0167] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization method, characterized in that, Includes the following steps: Based on the trained overcomplete dictionary, the timing clock deviation data of each underwater node is sparsely decomposed through the transmitting device of the deep-water heterogeneous network to obtain the original sparse vector. After the dynamic compressed observation data corresponding to the original sparse vector is converted into an optical signal, the transmitting device transmits the optical signal to the receiving device of the deep-water heterogeneous network via an umbilical cable; wherein, the dynamic compressed observation data is obtained by compressing and sampling the original sparse vector through a dynamic chaotic observation matrix generated by the timing clock deviation data. After the optical signal is demodulated into an observation vector by the receiving device, it is input into the trained deep unfolded network model, and the deep unfolded network model outputs a reconstructed sparse vector. The deep unfolded network model is based on a fixed-layer neural network architecture and is obtained by end-to-end training of each layer of the network using training pairs consisting of simulated sparse vectors and noisy chaotic observation vectors. The overcomplete dictionary is used to restore the reconstructed sparse vector into a clock deviation signal, and a clock synchronization calibration command is generated to achieve data synchronization of the corresponding underwater node.

2. The method according to claim 1, characterized in that, Before performing sparse decomposition on the timing clock offset data of each underwater node through the transmitting device of the deep-water heterogeneous network, the method includes the following steps: Based on a randomly generated initial dictionary, a sparse coefficient matrix is ​​calculated that minimizes the reconstruction error between the reconstructed sample matrix and the training sample set, while satisfying sparsity constraints. The training sample set includes time-series clock offset data of multiple underwater nodes in multiple time periods of the deep-water heterogeneous network. The reconstructed sample matrix is ​​obtained by multiplying the initial dictionary with the sparse coefficient matrix. Based on the non-zero coefficients of the sparse coefficient matrix, the sample index corresponding to the initial dictionary is determined, and the restricted residual matrix corresponding to the sample index is constructed. Based on the new atoms and coefficients obtained by performing singular value decomposition on the restricted residual matrix, the initial dictionary and sparse coefficient matrix are iteratively updated until the change in the reconstruction error is less than a preset threshold or the maximum number of iterations is reached, thus obtaining a trained overcomplete dictionary.

3. The method according to claim 2, characterized in that, Constructing the restricted residual matrix corresponding to the sample index includes the following steps: Based on the initial dictionary, a residual matrix corresponding to each column of atoms is constructed; wherein, the residual matrix is ​​constructed based on the difference between the training sample set and the corresponding reconstruction matrix; the reconstruction matrix is ​​the sum of multiplying the atoms in other columns except the current column of atoms with the corresponding coefficient row vectors in the sparse coefficient matrix; Based on the sample index, the corresponding restricted residual matrix is ​​selected from each of the residual matrices.

4. The method according to claim 2, characterized in that, Based on the new atoms and coefficients obtained by singular value decomposition of the restricted residual matrix, the initial dictionary and sparse coefficient matrix are iteratively updated, including the following steps: Singular value decomposition is performed on the restricted residual matrix to obtain the corresponding left singular vector matrix, diagonal matrix, and right singular matrix; Based on the left singular vector matrix, the diagonal matrix, and the right singular matrix, corresponding new atoms and new coefficients are generated; wherein, the new atom is the first column of the left singular vector matrix; and the new coefficient is the product of the first element of the diagonal matrix and the first column of the right singular matrix.

5. The method according to claim 1, characterized in that, Converting the dynamically compressed observation data corresponding to the original sparse vector into an optical signal includes the following steps: Based on the current timestamp of the time-series clock deviation data, a chaotic initial value is generated through a one-way irreversible numerical mapping operation; wherein the initial value is in the range of 0 to 1; The initial values ​​are input into the chaotic mapping equation for iterative calculation to generate a dynamic chaotic observation matrix that satisfies a preset length. Based on the dynamic chaotic observation matrix, the original sparse vector is compressed and sampled to obtain dynamic compressed observation data; Based on the mapping relationship between electrical signals and the phase difference of adjacent optical pulses, the dynamically compressed observation data is converted into optical signals.

6. The method according to claim 5, characterized in that, The initial values ​​are input into the chaotic mapping equation for iterative calculation to generate a dynamic chaotic observation matrix that satisfies a preset length, including the following steps: The initial value is input into the chaotic mapping equation for iterative calculation. The transient values ​​generated by the previous set number of iterations are discarded, and a pseudo-random sequence of a preset length is generated. The purified pseudo-random sequence is rearranged into an initial matrix of a preset dimension according to row priority or column priority rules. After normalizing the column vectors of the initial matrix, the column vectors are then optimized by vector projection decomposition to obtain the final dynamic chaotic observation matrix.

7. The method according to claim 1, characterized in that, Before inputting the observation vectors into the trained deep unfolded network model, the method further includes the following steps: A deep unfolded network model is constructed based on a neural network architecture with a fixed number of layers. The operation rules of the deep unfolded network model include: based on a learnable weight matrix, the received observation information is fused with the reconstruction results of the previous layer, and then the reconstruction results of each layer are obtained by soft thresholding. Based on the training pairs consisting of randomly generated simulated sparse vectors that satisfy sparsity constraints and noisy chaotic observation vectors, and combined with the loss function, end-to-end training is performed on each of the weight matrices and the threshold parameters corresponding to the soft threshold calculation to obtain the trained deep unfolded network model.

8. The method according to claim 7, characterized in that, The expression for the operation rule of the k-th layer of the deep unfolded network is: ； in, This represents the output of the (k+1)th layer of the network; This represents the output of the k-th layer network. This represents the observation information received by the depth-unfolded network; This represents the first weight matrix corresponding to the observation information of the k-th layer network; This represents the second weight matrix corresponding to the output of the k-th layer of the network; This represents the soft-threshold activation function value of the k-th layer network. , This represents the input variable for the soft-threshold activation function; Represents a symbolic function; This represents the threshold parameter of the k-th layer network; This represents the maximum value function.

9. The method according to claim 1, characterized in that, Using the overcomplete dictionary, the reconstructed sparse vector is restored to a clock offset signal, and a clock synchronization calibration command is generated to calibrate the clock of the corresponding underwater node, including the following steps: The reconstructed sparse vector is restored to a clock skew signal using the overcomplete dictionary; The clock deviation signal is compared with a deviation threshold. When the absolute value of the clock deviation signal is greater than the deviation threshold, a clock synchronization calibration command is generated. The clock synchronization calibration command includes: the identifier of the target underwater node and the calibration value. The clock synchronization calibration command is sent from the receiving device to the corresponding target underwater node via the underwater local area network to achieve data synchronization of the corresponding underwater node.

10. A spatiotemporally structured sparse optically coded deep-water heterogeneous network data synchronization system, characterized in that, include: The transmitting device of the deep-water heterogeneous network includes: a shore-based smart clock center, wherein the sparse coding module of the shore-based smart clock center is used to perform sparse decomposition on the time-series clock deviation data of each underwater node based on a trained overcomplete dictionary to obtain an original sparse vector, and to compress and sample the original sparse vector using a dynamic chaotic observation matrix generated by the time-series clock deviation data to obtain dynamically compressed observation data; the optical emission module of the shore-based smart clock center is used to convert the dynamically compressed observation data into an optical signal. An umbilical cable, connected to the optical transmitting module, is used to transmit the optical signal from the transmitting device to the receiving device of the deep-sea heterogeneous network. The receiving device of the deep-water heterogeneous network includes: an underwater intelligent routing module, wherein the optical receiving module of the underwater intelligent routing module is connected to the umbilical cable for receiving the optical signal and demodulating the optical signal to obtain an observation vector; the edge computing unit of the underwater intelligent routing module is used to construct a neural network architecture based on a fixed number of layers, and to train each layer of the network using a training method consisting of simulated sparse vectors and noisy chaotic observation vectors to generate a deep unfolded network model, and the deep unfolded network model calculates and reconstructs the sparse vector based on the observation vector, and recovers the clock deviation signal from the reconstructed sparse vector through the overcomplete dictionary; the synchronization control module of the underwater intelligent routing module is used to compare the clock deviation signal with a deviation threshold and generate a clock synchronization calibration command to realize the data synchronization of the corresponding underwater nodes.

Images