An Adaptive and Efficient Communication Method for Marine Multi-Parameter Precision Measurement Instruments

Abstract

This invention provides an adaptive and efficient communication method for marine multi-parameter precision measurement instruments, belonging to the field of adaptive and efficient communication technology for marine multi-parameter precision measurement instruments. This invention predicts the underwater acoustic channel state by constructing a multi-scale time-frequency attention long short-term memory transformer hybrid model, uses sparse Bayesian learning to complete compressed sensing channel estimation, utilizes an event-driven compressed sensing asynchronous sampling architecture to achieve low-power acquisition of multi-parameter signals, robustly fuses multi-source marine measurement data through an optimal transmission multi-platform data fusion algorithm, mitigates concurrent collisions in the underwater acoustic network with an artificial intelligence distributed dynamic time slot allocation protocol, and performs georegistration of the fused multi-parameter profile using a factor graph optimization method. This solves the problem of low communication efficiency caused by the inability to adaptively match pilot resources and equalization calculations due to the drastic time-varying state of the underwater acoustic communication channel.

Description

Technical Field

[0001] This invention belongs to the field of adaptive and efficient communication technology for marine multi-parameter precision measuring instruments, and specifically relates to an adaptive and efficient communication method for marine multi-parameter precision measuring instruments. Background Technology

[0002] Marine multi-parameter precision measurement instruments rely on underwater acoustic communication networks to remotely transmit measurement data. Underwater acoustic channels exhibit strong time-varying characteristics due to multipath effects, Doppler shift, and changes in the marine dynamic environment. Existing underwater acoustic communication systems typically employ a Turbo equalization scheme with fixed pilot density and a fixed number of iterations. Channel state estimation relies on traditional least squares or least mean square error algorithms, and multi-node networking depends on carrier sense multiple access collision avoidance protocols. While these schemes are applicable to near-shore scenarios with fixed nodes and relatively stable channels, in deep-sea multi-platform collaborative measurement scenarios, channel states change rapidly with internal wave passage, tidal modulation, and vehicle movement. A fixed pilot density wastes spectrum resources when the channel is good, but fails to provide sufficient estimation accuracy when the channel deteriorates. A fixed number of iterations leads to redundant calculations in simple channels and insufficient convergence in complex channels. In other words, existing technologies suffer from the technical problem of drastic time-varying underwater acoustic communication channel states, resulting in an inability to adaptively match pilot resources with equalization calculations and low communication efficiency. Summary of the Invention

[0003] In view of this, the present invention provides an adaptive and efficient communication method for marine multi-parameter precision measuring instruments, which can solve the technical problems in the prior art where the drastic time-varying state of the underwater acoustic communication channel leads to the inability to adaptively match pilot resources and equalization calculations, resulting in low communication efficiency.

[0004] This invention is implemented as follows: This invention provides an adaptive and efficient communication method for a marine multi-parameter precision measuring instrument, comprising the following steps:

[0005] Collect historical sequence of underwater acoustic channel impulse response, estimated Doppler frequency shift, node spacing, navigation speed, measured sea surface wind and wave data, and historical bit error rate sequence. Input the above data into a multi-scale time-frequency attention long short-term memory converter hybrid model and output a channel state prediction vector.

[0006] The channel quality comprehensive evaluation value is calculated by taking the multipath delay spread prediction value, Doppler spread prediction value and prediction bit error rate in the channel state prediction vector as inputs, and the pilot density level and the number of Turbo equalization iterations are adjusted according to the interval to which the channel quality comprehensive evaluation value belongs.

[0007] Pilots are configured at a given pilot density level. A sparse Bayesian learning channel estimation method is adopted. The spatial-temporal sparsity of the underwater acoustic channel multipath structure is utilized to complete the channel estimation using a compressed sensing framework. Turbo equalization is performed in conjunction with a given number of Turbo equalization iterations.

[0008] Using hydrodynamic anomaly detection signals as the trigger source, a high-speed sampling window is dynamically triggered. An event-driven compressed sensing asynchronous sampling architecture is used to complete the acquisition of multi-parameter signals. During non-trigger periods, the sigma delta modulator is switched to incremental working mode.

[0009] Argo buoy measurement data, autonomous underwater vehicle measurement data, and satellite remote sensing measurement data are input into the optimal transmission multi-platform data quality assessment and fusion algorithm to calculate the data quality differences between each platform. The optimal transmission distance is converted into a normalized fusion weight matrix, the Wasserstein centroid is solved, and the fused multi-parameter profile is output.

[0010] Each node takes the channel busy / idle history sequence and local queue depth as input, outputs time slot selection action through a lightweight Actor network, and executes a distributed dynamic time slot allocation protocol based on deep reinforcement learning. At the same time, it integrates sparse long-baseline underwater acoustic positioning observations, flow field prior constraints, and inertial estimation constraints, and uses a factor graph optimization method to batch correct historical trajectories within a sliding window, outputs the corrected geographic coordinates, and performs geographic registration on the fused multi-parameter profile.

[0011] Among them, the network front-end of the multi-scale time-frequency attention long short-term memory transformer hybrid model is equipped with a multi-resolution wavelet decomposition module, which decomposes the historical sequence of the underwater acoustic channel impulse response into sub-band sequences of multiple scales, and each sub-band sequence is fed into the parallel long short-term memory branch.

[0012] Among them, the output of the parallel long short-term memory branch is concatenated in the frequency domain and then fed into the improved Transformer encoder. The self-attention query matrix of the improved Transformer encoder is generated by conditionalizing hydrological parameters, forming a conditional attention mechanism.

[0013] Among them, the improved Transformer coding layer is set with a dynamic depth routing gate. The dynamic depth routing gate determines whether to skip the current coding layer and directly pass to the next coding layer at a certain time step based on the channel complexity score.

[0014] Among them, the multipath delay predicted by the ray acoustic model is injected as a physical prior position code into the improved position embedding layer of the Transformer encoder, realizing the fusion of physical prior and data-driven approaches.

[0015] Among them, the training of the multi-scale time-frequency attention long short-term memory transformer hybrid model adopts the mean square error loss function and Adam optimizer, the routing score of the dynamic deep routing gate adopts the auxiliary entropy regularization loss constraint, and the model is compressed by INT8 quantization after training.

[0016] Among them, the comprehensive channel quality assessment value The calculation formula is: ,in , , are dimensionless weighted coefficients and , , , For reference only.

[0017] Among them, the comprehensive channel quality assessment value When the channel quality is greater than or equal to the high-end threshold, it is considered poor. When the channel quality is within the medium threshold range, it is determined to be moderate. When the value is below the low threshold, the channel quality is considered good. Each threshold is determined by scanning and optimizing through multiple rounds of actual tests with the lowest bit error rate as the objective function.

[0018] Among them, the sparse Bayesian learning channel estimation method establishes a hierarchical sparse prior on the channel impulse response coefficients using a Bayesian framework, and iteratively infers the hyperparameters of each delay unit through the expectation-maximization algorithm to achieve sparse reconstruction.

[0019] The hydrodynamic anomaly detection threshold was determined through statistical analysis of the energy distribution of historical hydrodynamic signals and multiple field experiments. During non-triggering periods, the sigma delta modulator switches to incremental operating mode and outputs incremental codes only when the input signal changes.

[0020] Among them, the optimal transmission multi-platform data quality assessment and fusion algorithm estimates the joint distribution of kernel density multi-parameters for the current batch of data on each platform, and uses the Sinkhorn algorithm to regularize the parameters. Fast solution to the optimal transport problem, regularization parameter The value ranges from 0.01 to 0.1, and the specific value is determined through cross-validation experiments.

[0021] The Wasserstein centroid is solved by iterative Sinkhorn-Knopp algorithm until convergence, and the fused multi-parameter profile is obtained by sampling the Wasserstein centroid distribution.

[0022] The lightweight Actor network adopts a fully connected structure, using a weighted combination of the total network throughput and end-to-end latency as the reward signal. Each node is trained independently and cooperates implicitly through a shared reward structure.

[0023] Among them, the factor graph optimization method models the autonomous underwater vehicle navigation problem as a probabilistic graph composed of variable nodes and factor nodes. It performs batch posterior inference on the factor graph through a message passing algorithm, and performs batch optimization on all historical position status nodes within the window when sparse long-baseline underwater acoustic positioning observations arrive.

[0024] Among them, the multi-scale time-frequency attention long short-term memory transformer hybrid model has 4 sub-bands, 2 long short-term memory branch layers, and 32 hidden layer dimensions. The improved Transformer encoder contains 3 coding layers, 4 multi-head self-attention heads, and 64 feedforward hidden layer dimensions. The overall number of parameters is controlled within 50k, and the inference latency is less than 5ms. The hydrodynamic anomaly detection threshold is 1.5 to 3.0 times the average historical signal energy. The high-end threshold is 1.5, and the low-end threshold is 0.8. The lightweight Actor network has 64 hidden layer dimensions and approximately 8k parameters.

[0025] This invention employs a multi-scale time-frequency attention long short-term memory transformer hybrid model to predict the underwater acoustic channel state, and uses the prediction results to drive the adaptive adjustment of pilot density levels and the number of Turbo equalization iterations. This model simultaneously captures the dynamic patterns of two time scales—millisecond-level multipath jitter and minute-level tidal modulation—through multi-resolution wavelet decomposition. Combined with a conditional attention mechanism, hydrological parameters are introduced into the attention generation process, enabling the model to quickly adjust the region of interest based on physical priors when the channel structure changes, thereby outputting an accurate channel state prediction vector. The comprehensive channel quality assessment value calculated based on this prediction vector can preemptively increase pilot density and equalization iterations before channel degradation, and reduce pilot overhead and iteration rounds when the channel is good, achieving on-demand allocation of spectrum and computational resources. In summary, this invention solves the technical problem mentioned in the background art where the drastic time-varying channel state of underwater acoustic communication leads to the inability to adaptively match pilot resources and equalization calculations, resulting in low communication efficiency. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a time series plot of the comprehensive channel quality assessment values ​​during the experiment.

[0028] Figure 3 This is a comparison chart of the total network throughput under different numbers of nodes.

[0029] Figure 4 A comparison chart of the trajectory errors of the autonomous underwater vehicle before and after factor plot optimization. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0031] like Figure 1 The diagram shown is a flowchart of an adaptive and efficient communication method for a marine multi-parameter precision measuring instrument provided by the present invention. This method includes the following steps:

[0032] S01. Collect the historical sequence of the underwater acoustic channel impulse response, the estimated value of Doppler frequency shift, the node spacing, the navigation speed, the measured data of sea surface wind and waves, and the historical bit error rate sequence. Input the above data into the multi-scale time-frequency attention long short-term memory converter hybrid model and output the channel state prediction vector.

[0033] S02. Using the multipath delay spread prediction value, Doppler spread prediction value and prediction bit error rate in the channel state prediction vector output by S01 as input, calculate the comprehensive channel quality assessment value, and adjust the pilot density level and the number of Turbo equalization iterations according to the interval to which the comprehensive channel quality assessment value belongs.

[0034] S03. Configure pilots according to the pilot density level determined in S02, adopt the sparse Bayesian learning channel estimation method, utilize the spatiotemporal sparsity of the underwater acoustic channel multipath structure, complete the channel estimation with the compressed sensing framework, and perform Turbo equalization in combination with the number of Turbo equalization iterations determined in S02 to eliminate residual inter-symbol interference.

[0035] S04. Using the hydrodynamic anomaly detection signal as the trigger source, the high-speed sampling window is dynamically triggered. The event-driven compressed sensing asynchronous sampling architecture is used to complete the acquisition of multi-parameter signals. During non-triggering periods, the sigma delta modulator is switched to the incremental working mode to reduce power consumption during idle periods.

[0036] S05. Input the Argo buoy measurement data, autonomous underwater vehicle measurement data and satellite remote sensing measurement data into the optimal transmission multi-platform data quality assessment and fusion algorithm, calculate the optimal Wasserstein-2 transmission distance between the data of each platform, convert it into a normalized fusion weight matrix, solve for the Wasserstein centroid, and output the fused multi-parameter profile.

[0037] S06. Each node takes the channel busy / idle history sequence and local queue depth as input, outputs the time slot selection action through a lightweight Actor network, and executes a distributed dynamic time slot allocation protocol based on deep reinforcement learning. At the same time, it integrates sparse long baseline underwater acoustic positioning observations, flow field prior constraints, and inertial estimation constraints, and uses a factor graph optimization method to batch correct historical trajectories within a sliding window, outputs the corrected geographic coordinates, and performs geographic registration on the fused multi-parameter profile.

[0038] The specific structure of the multi-scale time-frequency attention long short-term memory transformer hybrid model is as follows: A multi-resolution wavelet decomposition module is set at the network front end to decompose the original underwater acoustic channel impulse response history sequence into sub-band sequences of four scales; the four sub-band sequences are respectively fed into four parallel long short-term memory branches, each branch having two layers and a hidden layer dimension of 32, used to capture the multipath evolution patterns at different time scales, such as millisecond-level multipath jitter and minute-level tidal modulation; the outputs of the four parallel long short-term memory branches are concatenated in the frequency domain and then fed into an improved Transformer encoder; the improved Transformer encoder contains three coding layers, each layer having four multi-head self-attention heads and a feedforward hidden layer dimension of 64; the self-attention query matrix is ​​generated conditionally from hydrological parameters represented by the thermocline depth, forming a... An attention mechanism is implemented to allow the attention head to explicitly focus on physically relevant time-frequency regions. A dynamic deep routing gate is set between the improved Transformer coding layers. This gate determines whether to skip the current coding layer and directly proceed to the next coding layer at a given time step based on the channel complexity score, reducing computational load during stable channels. Multipath delays predicted by the ray-acoustic model are injected as physical prior position codes into the position embedding layer of the improved Transformer encoder, achieving a fusion of physical priors and data-driven approaches. The network output layer is a linear mapping layer, outputting a channel state prediction vector. The dimension of this channel state prediction vector is consistent with the number of pilot subcarriers, and it includes multipath delay spread predictions, Doppler spread predictions, and prediction bit error rate. The overall parameter count is controlled within 50k, and the inference latency is less than 5kbps. .

[0039] The steps for establishing the training dataset for the multi-scale time-frequency attention long short-term memory transformer hybrid model specifically include: deploying underwater acoustic communication nodes multiple times in the experimental sea area, recording the historical sequences of underwater acoustic channel impulse responses under different seasons and sea conditions, with a collection duration covering 10... ~72 Simultaneously collect Doppler frequency shift estimates, node spacing, sailing speed, measured sea surface wind and waves, and historical bit error rate sequences; use the synchronously collected data as input and the subsequently measured channel state as label to construct supervised training sample pairs; perform multi-resolution wavelet decomposition preprocessing and normalization on the historical data, and divide the training set, validation set, and test set in an 8:1:1 ratio.

[0040] The specific steps for training the multi-scale time-frequency attention long short-term memory transformer hybrid model include: using the mean squared error loss function to measure the channel state prediction error; employing the Adam algorithm as the optimizer, with an initial learning rate of... The learning rate is decayed using a cosine annealing strategy; the batch size is 32, the training rounds are 200, and the early stopping threshold is that the validation set loss does not decrease for 20 consecutive rounds; the routing score of the dynamic deep routing gate is constrained by auxiliary entropy regularization loss to prevent routing degradation; after training, the model is compressed using INT8 quantization to reduce inference power consumption.

[0041] The multi-scale time-frequency attention long short-term memory transformer hybrid model simultaneously captures the dynamic patterns of two time scales in the underwater acoustic channel: short-term multipath jitter and long-term tidal modulation, through multi-resolution wavelet decomposition. The long short-term memory branch is responsible for time-series modeling, while the improved Transformer encoder is responsible for cross-time-step global dependency modeling. The conditional attention mechanism introduces hydrological parameters into the attention generation process, enabling the model to quickly adjust the region of interest based on physical priors when the channel structure changes. The dynamic deep routing gate automatically reduces the number of computation layers during stable channels, realizing on-demand allocation of computational resources. The physical prior position encoding embeds the ray acoustic model prediction results into the improved Transformer encoder, reducing the dependence of the data-driven model on a large number of labeled samples. The overall mechanism enables the model to accurately predict the underwater acoustic channel state on a low-power embedded platform, providing a basis for the adaptive adjustment of the pilot density level and the number of Turbo equalization iterations in S02.

[0042] The formula for calculating the comprehensive channel quality assessment value is as follows:

[0043] ;

[0044] in This is a dimensionless comprehensive evaluation value for channel quality. This is the predicted value for multipath delay spread, in units of... , For reference multipath delay spread, the unit is . , These are Doppler extended predictions, in units of , For reference to Doppler extension, the unit is... , To predict the bit error rate, For reference bit error rate, , , The coefficients are dimensionless weighted coefficients and The above reference values , , and weighting coefficients , , The results were determined through statistical analysis of multiple measured data in the experimental sea area and through multiple rounds of comparative experiments; when When the channel quality is deemed poor, the pilot density is adjusted to a high level, and the number of Turbo equalization iterations is set to 6-8. When the channel quality is determined to be moderate, the pilot density is adjusted to a medium level, and the number of Turbo equalization iterations is set to 3-5. When the channel quality is deemed good, the pilot density is adjusted to a low level, and the number of Turbo equalization iterations is set to 1 to 2. The thresholds of 0.8 and 1.5 are determined by scanning and optimizing the thresholds with the lowest bit error rate as the objective function through multiple rounds of field tests under different sea conditions in the experimental sea area.

[0045] The principle and technical effects of the sparse Bayesian learning channel estimation method are as follows: The multipath structure of the underwater acoustic channel exhibits sparse distribution characteristics in both the time delay domain and the spatial domain, meaning that the actual effective number of multipaths is much smaller than the number of resolution units corresponding to the maximum time delay spread. The sparse Bayesian learning method establishes a hierarchical sparse prior on the channel impulse response coefficients using a Bayesian framework. It iteratively infers the hyperparameters of each time delay unit through the expectation-maximization algorithm, automatically concentrating energy on the real multipath locations while compressing the coefficients of non-multipath locations to near zero, thus achieving sparse reconstruction. The compressed sensing framework utilizes sparsity to complete the construction of the channel measurement matrix with a pilot overhead far less than the number of pilots required for Nyquist sampling, reducing the spectrum resources occupied by pilots. When the number of multipaths is sparse, the method obtains an estimation result close to the real channel with fewer observations, providing a high-quality channel prior for Turbo equalization, enabling Turbo equalization to converge to a low bit error rate state within the number of iterations determined by S02, effectively balancing the contradiction between spectrum efficiency and communication reliability.

[0046] The Turbo equalization is an iterative receiver structure that drives the equalizer and the channel decoder through external information exchange. The equalizer outputs soft information as the input of the channel decoder, and the channel decoder outputs prior information to the equalizer. The two alternately iterate to gradually eliminate inter-symbol interference. As the number of iterations increases, the residual interference gradually decreases.

[0047] The principle and technical effects of the event-driven compressed sensing asynchronous sampling architecture are as follows: A low-power hydrodynamic anomaly detection circuit continuously monitors signal energy, triggering a high-speed sampling window only when the detected energy exceeds the hydrodynamic anomaly detection threshold. This hydrodynamic anomaly detection threshold is determined through statistical analysis of historical hydrodynamic signal energy distribution and multiple field experiments, ranging from 1.5 to 3.0 times the average historical signal energy. The specific value is determined through multiple rounds of experiments involving joint optimization of the false alarm rate and the missed detection rate. During non-triggering periods, the sigma delta modulator switches to incremental operating mode, outputting only the difference between adjacent samples to reduce digitization power consumption. Within the triggered high-speed sampling window, the compressed sensing framework undersamples the signal using a random measurement matrix, utilizing signal sparsity to reconstruct the complete waveform at the back end, compensating for the undersampling error introduced by the trigger delay. This mechanism enables the instrument to maintain sufficient time resolution during rapidly changing events such as internal wave passage, while reducing power consumption during stable periods, achieving a dynamic balance between sampling accuracy and power consumption.

[0048] The sigma delta modulator is an oversampling analog-to-digital converter structure that shifts quantization noise outside the signal band through noise shaping, achieving high effective bit resolution under low signal bandwidth conditions. The incremental operating mode means that the sigma delta modulator outputs incremental codes only when the input signal changes, and outputs the minimum number of bits during the static period, thereby reducing dynamic power consumption.

[0049] The principle and technical effect of the optimal transmission multi-platform data quality assessment and fusion algorithm are as follows: Multi-parameter measurement data from different platforms are treated as multi-dimensional probability measures. The Wasserstein-2 optimal transmission distance is used to quantify the physical consistency distance of data from different platforms in the multi-parameter joint distribution space. The smaller the Wasserstein-2 optimal transmission distance, the closer the probability distributions of the measurement results of the two platforms are, the higher the physical consistency, and the greater the corresponding fusion weight. The specific process is as follows: Estimate the kernel density multi-parameter joint distribution of the current batch of data for each platform; calculate the Wasserstein-2 optimal transmission distance pairwise for Argo buoy measurement data, autonomous underwater vehicle measurement data, and satellite remote sensing measurement data, and use the Sinkhorn algorithm to regularize the parameters. The regularization parameter is used to quickly solve the optimal transmission problem. The value ranges from 0.01 to 0.1, and the specific value is determined through cross-validation experiments on multiple sets of measured data. The Wasserstein-2 optimal transmission distance is converted into a normalized fusion weight matrix. The Wasserstein centroid is used as the optimal fusion result for multiple platforms, and the Wasserstein centroid is solved by iterative Sinkhorn-Knopp algorithm until convergence. The fused multi-parameter profile is obtained by sampling the Wasserstein centroid distribution. The algorithm does not require preset fixed weights for each platform. When a platform has a systematic deviation, its Wasserstein-2 optimal transmission distance increases, and the normalized fusion weight automatically decreases, realizing adaptive evaluation and robust fusion of data quality, and improving the estimation accuracy of key ocean parameters such as the depth of the mixed layer.

[0050] The Wasserstein-2 optimal transmission distance is the optimal transmission cost between two probability measures under the squared Euclidean cost, reflecting the geometric distance between the two probability distributions; the Sinkhorn algorithm transforms the original linear programming problem into a matrix scaling problem that can be solved iteratively by introducing an entropy regularization term; the Wasserstein centroid is the probability distribution that minimizes the sum of the weighted Wasserstein-2 optimal transmission distances from all input probability measures to the Wasserstein centroid, which is the optimal result of multi-platform distribution fusion.

[0051] The mixing layer depth is the lower boundary depth of a water layer in the upper ocean where temperature, salinity, and density are uniformly mixed vertically. It is a key ocean parameter characterizing the dynamic and thermal structure of the upper ocean.

[0052] The principle and technical effects of the distributed dynamic time slot allocation protocol based on deep reinforcement learning are as follows: The half-duplex characteristics and long propagation delay of the underwater acoustic channel cause the backoff mechanism of the traditional carrier sense multiple access collision avoidance protocol to fail in multi-node scenarios because the nodes cannot perceive the remote transmission status in time. In the deep reinforcement learning framework, each node takes the channel busy / idle history sequence and local queue depth as input states, outputs time slot selection actions through a lightweight Actor network, and uses a weighted combination of the total network throughput and end-to-end delay as the reward signal to train the lightweight Actor network online. The lightweight Actor network adopts a 2-layer fully connected structure with a hidden layer dimension of 64 and approximately 8k parameters. Each node trains independently but implicitly collaborates through a shared reward structure. When the node density changes, there is no need to redesign the protocol parameters, realizing distributed adaptive time slot management and effectively alleviating the problems of concurrent collisions and queuing delays.

[0053] The specific steps for acquiring training data for the lightweight Actor network include: setting up a multi-node underwater acoustic communication simulation environment and a field test environment in a laboratory tank and an actual sea area, respectively, recording the number of nodes as 2 to 16 and the propagation delay as 200. ~3000 Channel busy / idle history sequences and local queue depth sequences under certain conditions; large-scale reinforcement learning pre-training in a simulation environment, followed by online fine-tuning using measured data; the weight coefficients of total network throughput and end-to-end latency in the reward function are optimized and determined through multiple rounds of comparative experiments with the goal of maximizing total network throughput.

[0054] The principle and technical effect of the factor graph optimization method are as follows: The navigation problem of autonomous underwater vehicles is modeled as a probabilistic graph composed of variable nodes and factor nodes. The variable nodes represent the position and state at each time step, and the factor nodes correspond to inertial estimation constraints, flow field prior constraints, and sparse long baseline underwater acoustic positioning observation constraints, respectively. Batch posterior inference is performed on the factor graph through a message passing algorithm. The sliding window mechanism performs batch optimization on all historical position and state nodes within the window each time a sparse long baseline underwater acoustic positioning observation arrives. The absolute position information contained in the sparse long baseline underwater acoustic positioning observation is backpropagated to correct the accumulated error, and the positioning error is controlled within the meter level, thereby improving the spatial georegistration accuracy of the fused multi-parameter profile.

[0055] The sparse long-baseline underwater acoustic positioning observations are absolute position observations of autonomous underwater vehicles obtained by measuring the round-trip time delay of acoustic signals from multiple acoustic transponders deployed on the seabed. Because the number of transponders is limited and the update frequency is low, they are called sparse observations.

[0056] Optionally, the present invention also provides a computer-based method for forming an adaptive and efficient communication system for a marine multi-parameter precision measuring instrument. The computer is equipped with a readable storage medium that stores program instructions, which are used to execute the above-described method when the computer is run.

[0057] The specific implementation of step S01 is as follows: At the communication node of the marine multi-parameter precision measurement instrument, the historical sequence of the underwater acoustic channel impulse response, the estimated Doppler frequency shift, the node spacing, the navigation speed, the measured sea surface wind and wave data, and the historical bit error rate sequence are continuously collected. After multi-resolution wavelet decomposition preprocessing and normalization, the above input data is fed into a multi-scale time-frequency attention long short-term memory transformer hybrid model. The multi-resolution wavelet decomposition module at the network front end decomposes the historical sequence of the underwater acoustic channel impulse response into sub-band sequences of four scales, capturing the dynamic patterns of different frequency bands from millisecond-level multipath jitter to minute-level tidal modulation. Four parallel long short-term memory branches perform time-series modeling on each sub-band, and the outputs are spliced ​​in the frequency domain and fed into the improved Transformer encoder. The improved Transformer encoder uses a conditional attention mechanism to conditionally introduce hydrological parameters such as thermocline depth into the Query matrix generation, enabling the attention head to explicitly focus on physically relevant time-frequency regions. The dynamic depth routing gate determines whether to skip the current coding layer based on the channel complexity score, reducing computation during channel stability periods. The multipath delay predicted by the ray-acoustic model is injected into the location embedding layer as a physical prior location code. The network output layer outputs a channel state prediction vector through a linear mapping. The vector contains the multipath delay spread prediction value, the Doppler spread prediction value, and the prediction bit error rate, with the dimension consistent with the number of pilot subcarriers. The overall parameter count is controlled within 50k, and the inference latency is less than 5ms, meeting the real-time deployment requirements of embedded platforms.

[0058] The specific implementation of step S02 is as follows: using the multipath delay spread prediction value in the channel state prediction vector output by S01. Doppler extended prediction values and predicted bit error rate For input, according to the formula Calculate the dimensionless channel quality comprehensive evaluation value The reference value is... , , and weighting coefficients , , Statistical analysis and comparative experiments using multiple measured data from the experimental sea area determined that, .when When the channel quality is deemed poor, the pilot density is adjusted to a high level, and the number of Turbo equalization iterations is set to 6-8. When the channel quality is determined to be moderate, the pilot density is adjusted to a medium level, and the number of Turbo equalization iterations is set to 3-5. When the channel quality is deemed good, the pilot density is adjusted to a low level, and the number of Turbo equalization iterations is set to 1-2. The thresholds of 0.8 and 1.5 were determined through multiple rounds of experimental testing, with the lowest bit error rate as the objective function.

[0059] The specific implementation of step S03 is as follows: Pilot subcarriers are configured in the frequency domain according to the pilot density levels determined in S02, and the channel impulse response is reconstructed using a sparse Bayesian learning channel estimation method. The sparse Bayesian learning method establishes a hierarchical sparse prior for the channel impulse response coefficients using a Bayesian framework. It iteratively infers the hyperparameters of each delay unit through an expectation-maximization algorithm, automatically concentrating energy at the true multipath locations and compressing the coefficients at non-multipath locations to near zero. The compressed sensing framework utilizes the sparse distribution characteristics of the underwater acoustic channel in the delay and spatial domains, constructing a channel measurement matrix with far fewer observations than the number of pilots required for Nyquist sampling, thus completing the channel estimation. The obtained channel estimation result is used as prior information input to the Turbo equalizer. The equalizer and channel decoder mutually drive each other through external information exchange, performing iterations according to the number of iterations determined in S02, gradually eliminating residual inter-symbol interference, enabling the receiver to converge to a low bit error rate state within a finite number of iterations.

[0060] The specific implementation of step S04 is as follows: The low-power hydrodynamic anomaly detection circuit continuously monitors the input signal energy. The hydrodynamic anomaly detection threshold is set to 1.5 to 3.0 times the historical average signal energy, with the specific value determined through multiple rounds of experiments using joint optimization of the false alarm rate and the missed detection rate. When the detected signal energy exceeds the hydrodynamic anomaly detection threshold, a high-speed sampling window is triggered. An event-driven compressed sensing asynchronous sampling architecture is adopted, using a random measurement matrix to undersample the signal within the trigger window. The complete waveform is then reconstructed at the back end using signal sparsity to compensate for the undersampling error introduced by the trigger delay. During non-trigger periods, the sigma delta modulator switches to incremental operating mode, outputting only the difference between adjacent samples. It outputs the fewest bits when the input signal is stationary, effectively reducing dynamic power consumption. This mechanism enables the instrument to maintain sufficient time resolution during rapidly changing events such as internal wave transits and reduces energy consumption during stable periods.

[0061] The specific implementation of step S05 is as follows: Argo buoy measurement data, autonomous underwater vehicle measurement data, and satellite remote sensing measurement data are fed into the optimal transmission multi-platform data quality assessment and fusion algorithm. First, the kernel density multi-parameter joint distribution of the current batch of data for each platform is estimated. Then, the Wasserstein-2 optimal transmission distance is calculated pairwise for each of the three platforms, and the Sinkhorn algorithm is used to regularize the parameters. (Values ​​range from 0.01 to 0.1, with specific values ​​determined through cross-validation) This method rapidly solves the optimal transmission problem. A smaller Wasserstein-2 optimal transmission distance indicates a closer similarity in the probability distributions of the data from the two platforms, higher physical consistency, and a larger corresponding normalized fusion weight. Using the normalized fusion weight matrix as input, the Wasserstein centroid is solved iteratively using the Sinkhorn-Knopp algorithm until convergence. The fused multi-parameter profile is obtained by sampling the Wasserstein centroid distribution. When a platform exhibits systematic deviation, its Wasserstein-2 optimal transmission distance automatically increases, and the normalized fusion weight decreases accordingly, achieving adaptive assessment and robust fusion of data quality.

[0062] The specific implementation of step S06 is as follows: Each node takes the local channel busy / idle history sequence and local queue depth as input states, outputs time slot selection actions through a lightweight Actor network (using a 2-layer fully connected structure with a hidden layer dimension of 64), and updates network parameters online using a weighted combination of the total network throughput and end-to-end latency as the reward signal. Each node trains independently and implicitly collaborates through a shared reward structure, adapting to changes in node density without redesigning protocol parameters. Simultaneously, sparse long-baseline underwater acoustic positioning observations, flow field prior constraints, and inertial estimation constraints are fused to model the autonomous underwater vehicle navigation problem as a probabilistic graph composed of variable nodes and factor nodes. Variable nodes represent the position state at each time point, and factor nodes correspond to the three types of constraints. Batch posterior inference is performed using a message passing algorithm. Each time a sparse long-baseline underwater acoustic positioning observation arrives, batch optimization is performed on all historical position state nodes within the sliding window. Absolute position information is backpropagated to correct accumulated errors, and corrected geographic coordinates are output. Geographic registration is then performed on the fused multi-parameter profile.

[0063] It should be noted that the key technologies of this invention include: a multi-scale time-frequency attention long short-term memory transformer hybrid model decouples the channel sequence into sub-bands of different time scales through wavelet decomposition; the long short-term memory branch is responsible for the temporal modeling of each sub-band; an improved Transformer encoder captures global dependencies across time steps; a conditional attention mechanism introduces hydrophysical constraints; a dynamic depth routing gate allocates computational resources on demand; and physical prior position encoding reduces dependence on a large number of labeled samples. These multiple mechanisms work together to enable the model to achieve accurate channel state prediction on a low-power platform. The optimal transmission multi-platform data fusion algorithm treats measurement data as a multi-dimensional probability measure, quantifies the physical consistency between platforms using the Wasserstein-2 optimal transmission distance, and automatically reduces the fusion weight when a platform exhibits a systematic deviation, demonstrating significantly better robustness than a fixed-weight scheme. The synergy of these two key technologies is reflected in: accurate channel state prediction ensures the reliability of data transmission, while the robust multi-platform fusion algorithm ensures the quality of measurement data reaching the shore-based base. Together, they support the high-precision acquisition of multi-parameter oceanographic profiles.

[0064] It should be noted that this invention also solves the following technical problems: In existing technologies, multi-platform oceanographic measurement data fusion typically employs a fixed-weighted average scheme, which fails to detect data quality differences between platforms at different times and in different regions. When a platform exhibits systematic deviations due to sensor drift, biofouling, or cloud cover from satellite remote sensing, its biased data is incorporated into the fusion result with a fixed weight, leading to systematic errors in the estimation of key oceanographic parameters such as the mixed layer depth. Existing technologies suffer from the technical problem of failing to adaptively perceive the quality differences of multi-platform oceanographic measurement data, resulting in contamination of the fusion result by the biased platform. This invention, through an optimal transmission multi-platform data quality assessment and fusion algorithm, models the data of each platform as a multi-dimensional probability measure. The Wasserstein-2 optimal transmission distance quantitatively characterizes the physical consistency between platforms in the multi-parameter joint distribution space. Platforms with low physical consistency automatically receive lower normalized fusion weights. The Wasserstein centroid, as the optimal fusion result, provides a geometrically optimal compromise on the probability distribution of each platform, thereby achieving adaptive perception and robust fusion of multi-platform data quality without human intervention.

[0065] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems is that the time-varying nature of underwater acoustic channels is inherently predictable. Its changes are driven by physically governed ocean dynamic processes such as internal waves, tides, and wind waves. Therefore, historical channel impulse response sequences and synchronous hydrological parameters contain sufficient predictive information. The multi-scale time-frequency attention long short-term memory transformer hybrid model decomposes the channel sequence into sub-bands of different time scales through wavelet decomposition. The long short-term memory branch models the temporal evolution at each scale, improves the Transformer encoder to capture cross-time step global dependencies, and injects hydrological parameters such as thermocline depth into the attention weight generation process, ensuring that the model's prediction results are consistent with the actual physical change trends. After obtaining reliable predictions, the comprehensive channel quality assessment value integrates multipath delay spread, Doppler spread, and bit error rate normalization, providing a quantitative basis for switching pilot density and equalization iteration counts. This ensures that communication parameter configurations always match the current channel state, thereby maximizing spectral efficiency and energy efficiency while guaranteeing communication reliability.

[0066] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0067] The specific implementation method of step S01 is as follows.

[0068] Historical impulse response sequences, Doppler shift estimates, node spacing, navigation speed, measured sea surface wind and wave data, and historical bit error rate sequences of the underwater acoustic channel are collected and used as inputs to a multi-scale time-frequency attention long short-term memory transformer hybrid model. The network front-end performs multi-resolution wavelet decomposition on the historical impulse response sequences of the underwater acoustic channel, decomposing the original sequence into four sub-band sequences at four scales, denoted as follows: ,in Indicates scale index. For time step indexing. Four parallel long short-term memory branches perform temporal modeling on the four sub-band sequences, each branch has two layers and a hidden layer dimension of 32. The branch road in the The hidden state output of the time step is denoted as , It is a 32-dimensional real vector. The outputs of the four branches are concatenated in the frequency domain to obtain a 128-dimensional joint feature vector. Specifically, it is expressed as follows:

[0069] ;

[0070] In the formula, This is a 128-dimensional joint feature vector, and the semicolon indicates a vertical concatenation operation of the vectors. The signal is fed into the improved transformer encoder, which contains three coding layers, each with four multi-head self-attention heads and a feedforward hidden layer dimension of 64. In the conditional attention mechanism, the... The query matrix for layer attention is determined by the depth of the temperature-induced layer. Conditional generation, the calculation formula is expressed as follows:

[0071] ;

[0072] In the formula, For the first The layer condition query matrix is ​​a dimensionless real number matrix. For the first Layer query projection matrix, dimension is , is a dimensionless weight matrix obtained from model training. For the first The projection vector of the temperature jump condition has a dimension of , which is a dimensionless weight vector obtained from model training. This represents the model's feature dimension, with a value of 128. The thermocline depth is expressed in units of 1. , For reference thermocline depth, the unit is Experience value: 100 , The dimensionless normalized thermocline depth, This is the coding layer index, with values ​​of 1, 2, and 3. The dynamic deep routing gate is based on the channel complexity score. Decide whether to skip the current coding layer. From the current hidden state through a linear mapping, Function activation is obtained, The function maps real numbers to interval, when Below the threshold When this time step skips the current layer and proceeds directly to the next layer, The empirical value is 0.4. The ray acoustic model predicts the [number missing]th [unit missing]. Multipath delay As a physical prior location code injected into the location embedding layer, the location code vector Dimensions ,Depend on Generated according to standard sine and cosine encoding method, where Maximum latency spread, in units of , This is a multipath index. The network output layer is a linear mapping layer, outputting a channel state prediction vector. , dimension , For pilot subcarriers, Includes multipath delay spread predictions Doppler extended prediction values and predicted bit error rate During training, a mean squared error loss function is used with an additional route entropy regularization term. The formula for calculating the loss function is as follows:

[0073] ;

[0074] In the formula, This represents the dimensionless training loss value. For the first Channel state prediction values ​​for each pilot subcarrier, dimensionless. To correspond to the measured channel state label, dimensionless. is the entropy regularization coefficient, dimensionless, with an empirical value of 0.01. For routing score Information entropy, dimensionless. For the first The routing score for each time step is output by the dynamic deep routing gate, and its value range is [value range missing]. , This is the pilot subcarrier index. The training, validation, and test sets are divided in an 8:1:1 ratio. The optimizer uses the Adam algorithm, with an initial learning rate of... The batch size is 32, the number of training rounds is 200, and the model is compressed using INT8 quantization after training.

[0075] The specific implementation method of step S02 is as follows.

[0076] by , and The formula for calculating the overall channel quality assessment value, using the input as an example, is as follows:

[0077] ;

[0078] In the formula, This is a dimensionless comprehensive evaluation value for channel quality. This is the predicted value for multipath delay spread, in units of... , For reference multipath delay spread, the unit is . , These are Doppler extended predictions, in units of , For reference to Doppler extension, the unit is... , For predicting the bit error rate, dimensionless, For reference bit error rate, dimensionless. , , The coefficients are dimensionless weighted coefficients and satisfy the following conditions: The above parameters were all determined through statistical analysis of multiple measured data from the experimental sea area. When the channel quality is determined to be poor, the pilot density is adjusted to a high level, and the number of Turbo equalization iterations is set to 6-8. When the channel quality is determined to be moderate, the pilot density is adjusted to a medium level, and the number of Turbo equalization iterations is set to 3-5. If the channel quality is determined to be good, the pilot density is adjusted to a low level, and the number of Turbo equalization iterations is set to 1 to 2.

[0079] The specific implementation method of step S03 is as follows.

[0080] After configuring the pilots according to the pilot density levels determined in S02, a sparse Bayesian learning channel estimation method is employed. The underwater acoustic channel impulse response exhibits a sparse distribution in the time delay domain, with an effective multipath number... Much smaller than the maximum delay resolution unit number Sparse Bayesian learning methods for channel impulse response coefficient vectors (dimension is) A hierarchical sparse prior is established, and the hyperparameters of each delay unit are iteratively updated using the expectation-maximization algorithm. , For delay unit index, Control the first The prior variance of each time delay unit coefficient is automatically compressed to near zero for non-multipath position coefficients. After constructing the channel measurement matrix with far fewer pilots than required for Nyquist sampling, the compressed sensing framework performs Turbo equalization based on the number of Turbo equalization iterations determined by S02. The equalizer and channel decoder drive each other through external information exchange, gradually eliminating inter-symbol interference.

[0081] The specific implementation method of step S04 is as follows.

[0082] Low-power hydrodynamic anomaly detection circuit continuously monitors signal energy The unit is Hydrodynamic anomaly detection threshold From historical signal energy average Confirmed, unit is The range of values ​​is ~ The specific value was determined through multiple rounds of experiments using joint optimization of the false alarm rate and the missed detection rate. A high-speed sampling window is triggered during the active period. During non-triggering periods, the sigma delta modulator switches to incremental operating mode to reduce power consumption during idle periods. Within the trigger window, the compressed sensing framework undersamples the signal using a random measurement matrix, leveraging the signal sparsity to reconstruct the complete waveform at the back end.

[0083] The specific implementation method of step S05 is as follows.

[0084] Argo buoy measurement data, autonomous underwater vehicle measurement data, and satellite remote sensing measurement data are considered as multidimensional probability measures, denoted as follows: , , After estimating the joint distribution of kernel density and multiple parameters for the current batch of data on each platform, the optimal Wasserstein-2 transmission distance is calculated pairwise, and the Sinkhorn algorithm is used to regularize the parameters. Quick solution, The value ranges from 0.01 to 0.1, and was determined through cross-validation. (Two platforms) and The formula for calculating the dimensionless Wasserstein-2 optimal transmission distance is as follows:

[0085] ;

[0086] In the formula, For the platform With the platform The square of the dimensionless Wasserstein-2 optimal transmission distance between them. For joint probability distribution, For and For the set of all joint distributions of marginal distributions, and These are sample points in the measurement space of the two platforms. This represents the Euclidean distance between sample points, with units consistent with all measurement parameters. For reference distance scales, units are equal to Consistency, used for normalization, For the joint probability measure infinitesimal element, dimensionless, , This is the platform index, with values ​​of 1, 2, and 3. (For the platform...) The sum of its optimal Wasserstein-2 transmission distances to other platforms The calculation formula is expressed as follows:

[0087] ;

[0088] In the formula, For the platform The sum of dimensionless Wasserstein-2 optimal transmission distances with other platforms, dimensionless. Normalized fusion weights. The calculation formula is expressed as follows:

[0089] ;

[0090] In the formula, For the first The normalized fusion weights of each platform are dimensionless and satisfy... , For summation indexing. Based on the Wasserstein barycenter. The calculation formula, representing the optimal fusion result across multiple platforms, is as follows:

[0091] ;

[0092] In the formula, To ensure that all input platforms measure a fusion probability distribution that minimizes the sum of their weighted dimensionless Wasserstein-2 optimal transmission distances, an iterative Sinkhorn-Knopp algorithm is used until convergence. The fused multi-parameter profile was obtained from the sampling.

[0093] The specific implementation method of step S06 is as follows.

[0094] Each node uses the channel busy / idle history sequence With local queue depth For input, For length is A binary sequence, where element 0 indicates the channel is idle and element 1 indicates the channel is busy. This is the historical observation length, with an empirical value of 20. The unit is the number of data packets. The maximum queue depth is expressed in units of data packets, and the input state vector is given. The definition is as follows:

[0095] ;

[0096] In the formula, The input state vector is dimensionless. The queue depth is dimensionless and normalized. The time-slot selection action is output through a lightweight Actor network. The Actor network employs a two-layer fully connected structure with a hidden layer dimension of 64 and approximately 8k parameters. The reward signal is a weighted combination of the network's throughput and end-to-end latency. Each node is trained independently and implicitly collaborates through a shared reward structure. Simultaneously, sparse long-baseline underwater acoustic positioning observations are used... The flow field prior constraints and inertial estimation constraints are modeled as a factor graph, with variable nodes representing the position and state at each time step. The unit is The factor nodes correspond to inertial estimation constraints, flow field prior constraints, and sparse long-baseline underwater acoustic positioning observation constraints, respectively. This provides an index of observation times for long-baseline underwater acoustic positioning. Within a sliding window, each... Upon arrival, batch posterior inference is performed on all historical location status nodes within the window. Accumulated errors are corrected through message passing algorithm, and corrected geographic coordinates are output. Geographic registration is performed on the fused multi-parameter profile to control the positioning error within the meter level.

[0097] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: Technicians deployed an underwater acoustic communication and measurement network consisting of 8 nodes in a certain marginal sea area, including 3 autonomous underwater vehicles, 4 fixed mooring nodes and 1 surface buoy equipped with Argo buoy data relay function, to verify the method described in this invention throughout the entire process.

[0098] The test area featured significant seabed topography, with depths ranging from 200 to 850 meters and a thermocline depth of approximately 60 to 90 meters. The experiment lasted 72 hours, during which two internal wave events and one sea state (level 6) event were observed. In the channel state prediction phase, the input acquisition window of the multi-scale time-frequency attention long short-term memory transformer hybrid model covered the previous 60 minutes of the underwater acoustic channel impulse response history. Multi-resolution wavelet decomposition divided the original sequence into four scale sub-bands. The hidden layer dimension of each of the four parallel long short-term memory branches was 32. The improved Transformer encoder contained three coding layers, four multi-head self-attention heads, and a feedforward hidden layer dimension of 64. The overall parameter count was 48k, meeting the constraint of less than 50k parameters. The measured single inference latency on the embedded platform was 3.8ms.

[0099] During the comprehensive channel quality assessment phase, reference values ​​were determined based on statistical analysis of historical measured data from the experimental sea area. It takes 8ms. 2Hz for Weighting coefficients , , Comprehensive channel quality assessment value during the experiment Time series such as Figure 2 As shown, during the passage of internal waves... When the value rises significantly, exceeding the high-end threshold of 1.5, the system automatically adjusts the pilot density to a high level and the number of Turbo equalization iterations to 7; during the stable period... When the value drops below 0.8, the pilot density is switched to a low level, and the number of Turbo equalization iterations is reduced to 2, demonstrating the dynamic response capability of adaptive adjustment.

[0100] During the multi-parameter signal acquisition phase, the hydrodynamic anomaly detection threshold was set to 2.2 times the historical signal energy average. This value was determined through statistical analysis of historical data from the 6 hours prior to the experiment, with a false alarm rate not exceeding 5% and a false alarm rate not exceeding 8%, and was optimized through multiple rounds of field experiments. Both internal wave transit events were accurately triggered. Within the trigger window, high-speed sampling was performed using a compressed sensing framework to complete undersampling and back-end reconstruction. During non-trigger periods, the sigma delta modulator switched to incremental operating mode, significantly reducing digitization power consumption during idle periods compared to continuous high-speed sampling.

[0101] During the multi-platform data fusion phase, autonomous underwater vehicle (AUV) measurement data, Argo buoy relay data, and satellite remote sensing data were simultaneously integrated into the optimal transmission multi-platform data quality assessment and fusion algorithm. In the 38th hour of the experiment, AUV No. 1, equipped with a temperature, salinity, and depth sensor, experienced sensor drift, with its temperature measurement systematically exceeding the normal range by approximately 0.6°C. The optimal transmission algorithm detected a significant increase in the Wasserstein-2 optimal transmission distance between this platform and the other two platforms. Its normalized fusion weight automatically decreased from the normal level, effectively suppressing the impact of biased data on the fusion results. The regularization parameter of the Sinkhorn algorithm was also adjusted. The value was set to 0.05, determined through cross-validation experiments. The statistics of the optimal Wasserstein-2 transmission distance for each platform are shown in Table 1.

[0102] Table 1. Statistics on the optimal transmission distance of Wasserstein-2 between different platforms

[0103]

[0104] As shown in Table 1, after sensor drift occurred, the optimal Wasserstein-2 transmission distance between Autonomous Underwater Vehicle No. 1 and the other two platforms increased significantly, while the distance between the normal platforms remained stable, verifying the algorithm's ability to automatically identify the deviating platforms.

[0105] In the distributed dynamic time slot allocation phase, a lightweight Actor network with 8 nodes was pre-trained in a laboratory water tank simulation environment. The propagation delay range was set to 200–2500 ms, and the number of nodes was gradually increased from 2 to 8 for course-style training, followed by online fine-tuning using measured data. The reward function had a network throughput weighting coefficient of 0.7 and an end-to-end latency weighting coefficient of 0.3, which were optimized through multiple rounds of comparative experiments with the goal of maximizing network throughput. Figure 3As shown, the deep reinforcement learning distributed dynamic time slot allocation protocol maintains relatively stable overall network throughput as the number of nodes increases from 2 to 8, while the traditional carrier sense multiple access collision avoidance protocol experiences a rapid increase in collision rate and a significant decrease in throughput when the number of nodes exceeds 4, verifying the advantages of the protocol of this invention in multi-node scenarios.

[0106] During the georeferencing phase, four acoustic transponders were deployed on the seabed, and long-baseline underwater acoustic positioning observations were conducted approximately three times per hour. The sliding window length for factor graph optimization was set to 10 minutes, and batch optimization was performed on all historical location status nodes within the window each time a sparse long-baseline underwater acoustic positioning observation arrived. For example... Figure 4 As shown, the accumulated error of the autonomous underwater vehicle trajectory optimized by the factor map is significantly reduced compared with the trajectory calculated solely by inertia, and the positioning error converges to the meter level. The spatial geographic registration accuracy of the fused multi-parameter profile is effectively improved.

[0107] After the experiment, the estimation results of the mixing layer depth in the fused multi-parameter profile were evaluated, using data from an independent shipborne acoustic Doppler current profiler as a control. Table 2 shows the time-by-time comparison of the mixing layer depth between the fused results and the control data, listing the comparative data for representative time intervals.

[0108] Table 2 Comparison of Mixed Layer Depth Estimation Results

[0109]

[0110] As shown in Table 2, at the moment when sensor drift occurs at 38h, the estimation results relying solely on single-platform data show significant deviations, while the multi-platform fusion results of this invention remain consistent with the independent control values, verifying the robustness of the algorithm.

[0111] During the entire 72-hour experiment, the statistical results of the cumulative number of triggers and the total duration of the trigger window for the event-driven compressed sensing asynchronous sampling architecture show that during the triggering period, high-resolution sampling fully captured fast-changing events such as internal wave transit. During the non-triggering period, the incremental working mode of the sigma delta modulator effectively reduced the number of output bits, and the digitization power consumption during the idle period was significantly lower than that of the continuous high-speed sampling mode.

[0112] The technological advancements of this invention compared to traditional methods are reflected in the following aspects. First, traditional underwater acoustic communication uses fixed pilot density and fixed equalization iterations, which cannot perceive the time-varying nature of channel states. This results in redundant overhead when the channel is good and insufficient estimation accuracy when the channel deteriorates. This invention uses a multi-scale time-frequency attention long short-term memory converter hybrid model to predict the channel state change trend in advance, and transforms the prediction results into an adaptive configuration of pilot density and equalization parameters. In principle, this achieves dynamic matching between communication parameters and channel states, overcoming the inherent limitations of static configuration. Second, traditional multi-platform data fusion relies on fixed weights and cannot perceive the dynamic changes in platform data quality. This invention uses the Wasserstein-2 optimal transmission distance as a geometric measure of data quality. Its physical meaning is clear and it is sensitive to changes in probability distribution. When the data distribution of a certain platform deviates from that of other platforms, the fusion weight automatically decreases. This achieves adaptive evaluation of multi-platform data quality at the probability distribution level, fundamentally avoiding the systematic contamination of the fusion results by biased data and improving the reliability of key oceanographic parameter estimation.

[0113] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0114] Table 3. Variable Explanation Table (Part 1)

[0115]

[0116] Table 4. Variable Explanation Table (Part Two)

[0117]

[0118] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive and efficient communication method for a marine multi-parameter precision measuring instrument, characterized in that, Includes the following steps: Collect historical sequence of underwater acoustic channel impulse response, estimated Doppler frequency shift, node spacing, navigation speed, measured sea surface wind and wave data, and historical bit error rate sequence. Input the above data into a multi-scale time-frequency attention long short-term memory converter hybrid model and output a channel state prediction vector. The channel quality comprehensive evaluation value is calculated by taking the multipath delay spread prediction value, Doppler spread prediction value and prediction bit error rate in the channel state prediction vector as inputs, and the pilot density level and the number of Turbo equalization iterations are adjusted according to the interval to which the channel quality comprehensive evaluation value belongs. Pilots are configured at a given pilot density level. A sparse Bayesian learning channel estimation method is adopted. The spatial-temporal sparsity of the underwater acoustic channel multipath structure is utilized to complete the channel estimation using a compressed sensing framework. Turbo equalization is performed in conjunction with a given number of Turbo equalization iterations. Using hydrodynamic anomaly detection signals as the trigger source, a high-speed sampling window is dynamically triggered. An event-driven compressed sensing asynchronous sampling architecture is used to complete the acquisition of multi-parameter signals. During non-trigger periods, the sigma delta modulator is switched to incremental working mode. Argo buoy measurement data, autonomous underwater vehicle measurement data, and satellite remote sensing measurement data are input into the optimal transmission multi-platform data quality assessment and fusion algorithm to calculate the data quality differences between each platform. The optimal transmission distance is converted into a normalized fusion weight matrix, the Wasserstein centroid is solved, and the fused multi-parameter profile is output. Each node takes the channel busy / idle history sequence and local queue depth as input, outputs time slot selection action through a lightweight Actor network, and executes a distributed dynamic time slot allocation protocol based on deep reinforcement learning. At the same time, it integrates sparse long-baseline underwater acoustic positioning observations, flow field prior constraints, and inertial estimation constraints, and uses a factor graph optimization method to batch correct historical trajectories within a sliding window, outputs the corrected geographic coordinates, and performs geographic registration on the fused multi-parameter profile.

2. The adaptive and efficient communication method for a marine multi-parameter precision measuring instrument according to claim 1, characterized in that, The network front-end of the multi-scale time-frequency attention long short-term memory transformer hybrid model is equipped with a multi-resolution wavelet decomposition module, which decomposes the historical sequence of the underwater acoustic channel impulse response into sub-band sequences of multiple scales, and each sub-band sequence is fed into the parallel long short-term memory branch.

3. The adaptive high-efficiency communication method for marine multi-parameter precision measuring instruments according to claim 2, characterized in that, The output of the parallel long short-term memory branch is concatenated in the frequency domain and then fed into the improved Transformer encoder. The self-attention query matrix of the improved Transformer encoder is generated by conditionalizing hydrological parameters, forming a conditional attention mechanism.

4. The adaptive high-efficiency communication method for marine multi-parameter precision measuring instruments according to claim 3, characterized in that, An improved version of the Transformer coding layer is set up with a dynamic depth routing gate. The dynamic depth routing gate determines whether to skip the current coding layer and directly pass to the next coding layer at a certain time step based on the channel complexity score.

5. The adaptive high-efficiency communication method for marine multi-parameter precision measuring instruments according to claim 4, characterized in that, The multipath delay predicted by the ray acoustic model is injected as a physical prior position code into the improved position embedding layer of the Transformer encoder, realizing the fusion of physical prior and data-driven approaches.

6. The adaptive high-efficiency communication method for a marine multi-parameter precision measuring instrument according to claim 5, characterized in that, The training of the multi-scale time-frequency attention long short-term memory transformer hybrid model uses the mean squared error loss function and the Adam optimizer. The routing score of the dynamic deep routing gate is constrained by the auxiliary entropy regularization loss. After training, the model is compressed by INT8 quantization.

7. The adaptive high-efficiency communication method for marine multi-parameter precision measuring instruments according to claim 6, characterized in that, The sparse Bayesian learning channel estimation method establishes a hierarchical sparse prior on the channel impulse response coefficients using a Bayesian framework, and iteratively infers the hyperparameters of each delay unit through the expectation-maximization algorithm to achieve sparse reconstruction.

8. The adaptive high-efficiency communication method for a marine multi-parameter precision measuring instrument according to claim 7, characterized in that, The hydrodynamic anomaly detection threshold was determined through statistical analysis of the energy distribution of historical hydrodynamic signals and multiple field experiments. During non-triggering periods, the sigma delta modulator switches to incremental operating mode and outputs incremental codes only when the input signal changes.

9. The adaptive high-efficiency communication method for a marine multi-parameter precision measuring instrument according to claim 8, characterized in that, The lightweight Actor network adopts a fully connected structure, using a weighted combination of the network throughput and end-to-end latency as the reward signal. Each node is trained independently and cooperates implicitly through a shared reward structure.

10. The adaptive and efficient communication method for a marine multi-parameter precision measuring instrument according to claim 9, characterized in that, The factor graph optimization method models the autonomous underwater vehicle navigation problem as a probabilistic graph composed of variable nodes and factor nodes. It performs batch posterior inference on the factor graph through a message passing algorithm and performs batch optimization on all historical position status nodes within the window when sparse long-baseline underwater acoustic positioning observations arrive.

Images